Skip to content

feat(bwopenchamber): add bwopenchamber and bwopencode2 --serve-open - #43

Open
Stuart B. Wilkins (stuwilkins) wants to merge 54 commits into
NSLS2:mainfrom
stuwilkins:add_bwopenchamber
Open

Stuart B. Wilkins (stuwilkins) wants to merge 54 commits into
NSLS2:mainfrom
stuwilkins:add_bwopenchamber

Conversation

@stuwilkins

@stuwilkins Stuart B. Wilkins (stuwilkins) commented Nov 11, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Adds bwopenchamber, which runs OpenChamber (a web UI for OpenCode) in the same bubblewrap sandbox as its OpenCode, so the terminal, git, worktrees and every tool call stay sandboxed and only the browser is outside, behind an ssh tunnel.

Also changes bwopencode2 --serve (from #37) so it is easier to sign in to: one block with the ssh -L command, a one-time sign-in link and a google-chrome --app=... command, a new link on Enter, and --serve-open, which opens the signed-in web UI in a browser on the same host. These are six commits at the end of the branch and are independent of bwopenchamber.

This depends on #37 and should not merge before it. It uses that PR's private network namespace and publish address, so until #37 merges this PR also shows #37's commits. Review only the 17 commits after #37's tip e8738f0 (git log e8738f0.. lists them): eleven for bwopenchamber and its sign-in, database-lock and data-directory fixes, and six for bwopencode2 --serve-open and the sign-in block (feat(bwopencode2): add --serve-open and a sign-in block for --serve, fix(bwopencode2): do not hand the --serve-open link to xdg-open without a browser, docs(bwopencode2): record that only a browser signs in with the --serve link, fix(bwopencode2): keep the launcher's descriptors from the --serve-open browser, feat(bwopencode2): make --serve-open work without --serve, and the cleaned-up --serve-open terminal output). Once #37 merges, the diff here shrinks to those.

Description

Why a launcher of its own, and not bwopencode2 --serve behind an OpenChamber on another machine. OpenChamber is not only a screen. Its server runs the browser terminal, git and worktree commands, project start commands and scheduled tasks, and it gives the agent tools that call back into it. Run outside the sandbox in front of a sandboxed OpenCode, all of that would run unsandboxed. So the whole server runs inside one sandbox.

What it does

  • Sources bin/bwopencode2 (it now runs main only when executed, and finds its library from its own file) for the OpenCode 2 mounts, environment, database guard and the --serve port pool, so a fix made there applies to both. Moving that shared code into lib/ is deliberately left for after Feat/bwrap private netns #37 lands, because bwopencode2 is being rewritten there.
  • Takes its port from the same pool and claim directory as --serve (24700-24709), published on the user's 127.H1.H2.1; one ssh_config block (--ssh-config) covers both. The server listens on 0.0.0.0 inside the private network, as --serve does.
  • Random UI password for each session (or BWOPENCODE2_SERVE_PASSWORD), passed only in the sandbox environment.
  • A random sign-in signing key for each session (OPENCODE_JWT_SECRET). Without it OpenChamber 2.7.0 writes a key to its data directory and reuses it, so a browser stayed signed in across restarts and password changes, and the agent could read the file.
  • The policy is a read-only /etc/openchamber/policy.json that the launcher generates and mounts from a descriptor inside the sandbox (the host's file is not visible there): enterprise mode (no tunnels, public relay, classification service, adding providers in the app), the installed OpenCode 2 pinned, allowNetworkAccess: true for the private network, no extensions. It is a file, not environment variables, because without one a saved opencodeBinary in OpenChamber's settings.json (which the agent can write) overrides OPENCODE_BINARY; only a policy file outranks it. Copilot's review found this: a fake OpenCode named in settings.json ran with the environment alone, and did not run with the file.
  • No host OPENCODE_* or OPENCHAMBER_* variable reaches the server, and no OpenChamber argument is passed through.
  • Finds a Node.js that is new enough, not just the first on PATH: BWOPENCHAMBER_NODE if set, else the node on PATH if it is 24.14 or newer, else the newest node-24 / nodejs24 in /usr/bin and /usr/local/bin that is (RHEL 10 keeps nodejs 22 as the default and nodejs24 installs node-24 beside it). With no node on PATH it takes the newest versioned one (a bug found in Copilot's review: it first took the first one found). When none qualifies the error lists what it found with versions and what to do. Checked on the real host with no stubs: it lists /usr/bin/node and /usr/bin/node-22 (both 22.23.2), and BWOPENCHAMBER_NODE set to a pixi Node 24.22.1 launches. On babelfish (RHEL 10.2) the dry run, with no override, uses /usr/bin/node-24 (24.21.0, from nodejs24, installed next to the default 22.23.2), and a real launch started the server, the pinned OpenCode and the browser terminal's node-pty. The matching install is NSLS2/ansible#5064.
  • The data directory is bound whole and read-write, so it is checked in bind mode: a symlink to a directory that contains a blocked path (~/.config/gh) is refused, and the resolved path is what is mounted (found in Copilot's review; _check_dir_safe had accepted it).
  • OpenChamber's data directory is ~/.local/share/bwopenchamber, apart from any OpenChamber run outside a sandbox. Its run/ directory is an empty tmpfs, because OpenChamber writes the UI password there in plain text (checked with 2.7.0).

bwopencode2 --serve: sign-in block and --serve-open

  • When the server answers, the wrapper inside the sandbox prints one block: the ssh -L command and the ssh_config alias, the one-time sign-in link for http://127.0.0.1:PORT, the same link as a google-chrome --app=... command (an app window without tabs), the link for this host and the password. A link works once and expires after 5 minutes and an ssh tunnel can take longer to set up, so Enter in the terminal asks the server for a new one. The wrapper only reads Enter when it has a terminal and stops at end of input. A backslash or double quote in a fixed password is now escaped for curl's config file.
  • --serve-open is --serve plus a browser on the same host, so it works by itself (--serve --serve-open means the same) and --serve alone opens nothing. It starts a helper on the host, outside the sandbox, just before the sandbox replaces the launcher. The helper waits for the server, asks it for a one-time link of its own and starts the browser detached: google-chrome --app=LINK (or google-chrome-stable, chromium, chromium-browser), else firefox, else xdg-open but only when xdg-mime names a default web browser for http (the scheme of the link). BWOPENCODE2_SERVE_BROWSER names another one. With no desktop session (DISPLAY and WAYLAND_DISPLAY unset) or no browser it opens nothing and says so; --dry-run opens nothing.
  • The --serve-open terminal output is short, since it is for the host you are on: one line, OpenCode web UI on ADDR:PORT, opened in google-chrome., or, if no browser opened (none found, no desktop session, the browser exited with an error within 8 seconds, the server never answered), a fresh one-time link to open by hand, how to get another (Enter), and the password for "Use password" (only there). --serve output is unchanged.
  • The helper closes the launcher's descriptors before starting the browser. Found in Copilot's review of this PR and reproduced: the browser inherited the database session lock and held it as long as its window was open, so the next launch on the same data directory would have been refused. A test fails without the fix.
  • Why xdg-open is guarded: on a workstation with no browser, xdg-open handed the link to the GNOME text editor, which showed "Could Not Open File, HTTP Client Error: Unauthorized".
  • Only a browser signs in. The server redirects to / and sets the cookie only for Accept: text/html; curl or the text editor get a JSON token and use up the code (checked with OpenCode 2.2.7).

Known limit (documented in docs/bw-wrappers.md). OpenChamber puts its credentials into the environment of the OpenCode it starts, so the agent's shell has OPENCHAMBER_UI_PASSWORD, OPENCODE_SERVER_PASSWORD, OPENCODE_PASSWORD and the new OPENCODE_JWT_SECRET (checked in a running session). With them an agent could sign in to the UI (or sign a cookie itself) or call its own OpenCode server and so change settings or answer its own permission prompts. Nothing leaves the sandbox. The launcher cannot remove them, and hiding them only from the agent's shell would not help, because a process of the same user can read the server's environment from /proc. The agent is told not to use them; that is an instruction, not a control.

Open questions for reviewers (1, 3 and 6 especially for Thomas A Caswell (@tacaswell), as the author of #37)

  1. Sourcing bwopencode2 is the smallest change but couples the two launchers. Is a follow-up PR that moves the shared code into lib/ after Feat/bwrap private netns #37 the right way to untangle it?

  2. Sharing the --serve port pool and claim directory means one alias covers both. Any objection to the two sharing it?

  3. --share-net is still accepted. The server then binds the user's publish address on the host network, and OpenCode's own port is on the host's loopback behind its own random password. Keep it as the escape hatch for hosts without pasta or slirp4netns, or refuse it for OpenChamber?

  4. The published 127.H1.H2.1 address is not access control (the Feat/bwrap private netns #37 docs say so); another local user can connect, and the UI password keeps them out. Is that acceptable for a UI that gives a terminal, or should this wait for per-user access control on published ports?

  5. OpenChamber's own outbound traffic (update checks, GitHub) goes through the private network and is logged, but nothing filters it.

  6. On a real launch ss -ltnp shows the published socket owned by platform-python (the network log's helper) and not by pasta. Is that expected?

  7. Copilot's review summary of the second push says pasta startup "releases the database session lock", with no detail. Checked on a real launch (fuser on the lock file while a bwopencode2 session ran on pasta): the lock is held for the whole session and released after it ends, but only by the network log's helper processes (a forked copy of the launcher shell and its condenser), which inherited the descriptor before the launcher execed into pasta; pasta, bwrap and OpenCode do not hold it. By reading the code, with no log tool (neither systemd-cat nor logger) there is no helper and the lock would be released at the exec; that is not tested. It is Feat/bwrap private netns #37's launch code and affects bwopencode2 too, so the fix belongs there (I've asked Thomas A Caswell (@tacaswell) on Feat/bwrap private netns #37); the new paragraph in docs/bw-wrappers.md records what was found.

    Separately, Copilot's fourth review found a real bug with --share-net, and I confirmed it: oc2db_guard put the session lock on bash's first free descriptor (10 or 11), and the launcher's descriptor relay hands the sandbox its data on descriptors 10 and up, opening them over whatever is there. The relay closed the lock, so with bwopenchamber (policy 10, context 11) the lock was released while the sandbox ran; bwopencode2 (context 10, lock 11) kept it by luck. Fixed in lib/opencode2-db-lock.sh: the lock is now on a descriptor from 200 up. A lock on 200 stayed held through the real relay and a real bwrap. Two new checks in tests/test-opencode2-db-guard.sh (82 now) fail on the old code and pass on the new. With pasta, pasta closes the descriptor before the relay runs, so the relay never meets it there and the position described above is unchanged.

  8. --serve-open puts the one-time link on the browser's command line, so another user of the host can read it with ps for the moment the browser starts and sign in first (30-day cookie). The published address already accepts every local user, with the password as the only protection, and the link works once and expires in 5 minutes. A local redirect file would avoid it but could not be tested here. Acceptable, or should it wait for that? (docs/bw-wrappers.md says not to use it on a host where others log in.)

  9. The weak directory check described above (_check_dir_safe on a directory bound whole, which accepts a symlink to a directory containing ~/.config/gh) is on main in bwopencode2 (data, cache, TUI state) and, by reading, in bwclaude, bwcodex, bwcopilot and bwopencode. Reproduced for bwopencode2 and fixed here only for bwopenchamber; a comment on Feat/bwrap private netns #37 has the details and the suggested _bwrap_canonicalize_safe fix.

  10. Not in this PR: the ansible install (pinned @openchamber/web, Node 24.14 or newer), and OpenChamber is not on the approved-tools list, so DSSI approval is needed before any rollout.

Testing

Tested on a workstation with a test install of OpenChamber 2.7.0 (not through ansible), with OpenCode 2.2.7.

A real launch (bwopenchamber --port 24700, pasta):

  • The host has one new listener, 127.H1.H2.1:24700, and nothing on 3000 or on OpenCode's port. Without a password the API gives 401, the same port on 127.0.0.1 is refused, and OpenCode's own port is refused from the host.
  • A browser signed in through an ssh -L tunnel and an agent session worked through the model gateway.
  • OpenChamber's browser terminal runs inside the sandbox: BWRAP_SANDBOX is bwopenchamber, ~/.ssh does not exist, ssh is masked and BWRAP_NET_BACKEND is pasta.
  • After Ctrl-C the OpenCode that OpenChamber started is gone.
  • The database guard refused to start next to an older OpenCode build, as designed.
  • The sign-in key was checked against the real server in a private network namespace: without a key a cookie survives a restart and a new password and jwt-secret is written; with a new key per run the cookie is refused after a restart (401), signing in again works, and no file is written.

Tests and lint

  • tests/test-bwopenchamber.sh: 104 checks through --dry-run (argument refusals, a symlinked data directory (refused when it leads to a directory containing .config/gh), the environment and mounts built, the generated policy file, hostile host variables dropped, password and sign-in key handling, Node and install checks). tests/bwopenchamber-harness accepts the development machine's bwrap, and stubs the port probe, so a running session holding the first port cannot change the result (Copilot found that it did; checked that without the stub the same scenario gives the next port). The library's owner check is not changed for real runs. Two checks in the first version could not fail, because --dry-run prints variables named PASSWORD or SECRET as REDACTED; they now check the names.
  • tests/test-opencode2-db-guard.sh: 82 of 82 pass. tests/test_serve_timeout.py passes. bwopencode2 still runs main when executed directly and through a symlink.
  • pixi run -e dev lint on all files passes (every hook, including actionlint and zizmor), and the commit hook passes on the staged files.

--serve-open and the sign-in block

  • tests/test_serve_signin.py (25 tests, with tests/test_serve_timeout.py 73 pass in the whole suite): the in-sandbox wrapper against a fake server (the password reaches it through curl, including a quote and a backslash; a new code on Enter through a pseudo-terminal; no spinning at end of input; a failed pairing) and the host helper against a stub browser (--app= link, plain link, login page when no link came, a launcher that has gone, a helper that survives the launcher's exec, nothing on --dry-run, no password in the browser's arguments).
  • tests/test-bwopencode2-serve.sh: 57 checks through --dry-run (arguments, --serve-open by itself and with --serve, browser choice, no desktop session, no browser installed, a registered default browser, firefox). tests/bwopencode2-harness accepts the development machine's bwrap and stubs the port probe, as for bwopenchamber.
  • A real launch on babelfish (RHEL 10.2, Wayland) with a stand-in browser script: the helper started it once the server answered, with the right link; a browser-style fetch of the link gave 302 to / and the login cookie (value not logged). With no browser installed it opens nothing and says so. After installing Google Chrome, --serve-open works there.

Not tested: git and worktrees from OpenChamber, an install through ansible on a managed host, Enter for a new link inside the real sandbox's terminal (only through a pseudo-terminal), and --serve-open over ssh -X.

Checklist

  • Edit the sections above with relevant information
  • Update in-repo documentation, if applicable
  • Run pixi run -e dev lint locally (or wait for CI) — see .github/workflows/lint.yml
  • Test the changes on a representative managed host (or runansible deployment)

Assisted-by: opencode:claude-sonnet-5-5

Runs opencode2 serve inside the sandbox on 127.0.0.1, so tools stay sandboxed, with one server per invocation on the first free port of a fixed range (24700-24709 by default) so one ssh_config block covers every session. The password is random for each session unless BWOPENCODE2_SERVE_PASSWORD is set, and reaches the server only as OPENCODE_PASSWORD. Ports are claimed with a host-side file so two launchers cannot pick the same one. --serve-ssh-config prints the matching ssh_config block.

Assisted-by: opencode:claude-sonnet-5-5
The web UI opens on a page that asks for a pairing link or the password, and opencode pair points at the shared service, which the wrapper refuses. Inside the sandbox a small wrapper now starts the server, waits for it, asks it for a pairing code and prints the link, so the browser can sign in without typing the password. The password reaches curl on stdin, never on a command line. Use password still works after the link expires.

Assisted-by: opencode:claude-sonnet-5-5
pasta, which will start bwrap in a private network namespace, closes
every inherited descriptor above 2, so the --file and --ro-bind-data
descriptors opened by the launcher would never reach bwrap. The
launcher now only reserves descriptor numbers and records their source
(a host file, or generated text carried in the environment); a fixed
bash relay opens them at those numbers and execs bwrap.

bwcodex no longer stages temporary copies of its config and daemon
settings, and --dry-run shows real descriptor numbers instead of
placeholders.

Assisted-by: opencode:claude-opus-5-5
pasta (package passt) now starts each sandbox in its own network
namespace. Outbound connections still leave from the host's address,
but the host's loopback services are unreachable and the sandbox's
listeners are invisible to the host. There is no fallback to the host
network: without a working pasta the launchers exit 3.

Crossing the boundary is a per-port, per-launch opt-in:
- --publish PORT[:SANDBOXPORT] binds a per-user, per-session host
  address 127.H1.H2.S (--publish-addr overrides it, e.g. for OAuth);
- --host-port PORT[:SANDBOXPORT] makes the host's 127.0.0.1:PORT
  reachable as host.internal, with a deny list (CUPS, Docker, X11,
  Tailscale, the OpenCode 2 service);
- --net-log FILE records every flow.

pasta maps the caller to uid 0 in its user namespace, so bwrap now runs
the agent as the real uid in a nested one (--unshare-user --uid --gid);
without it the agent would be root with a full capability set there.
bwrap also restores the real hostname that pasta replaces.
/etc/resolv.conf and /etc/hosts are generated, the context file gains
a Network section, and BWRAP_NET_* describe the forwards.

This breaks --serve, will fix later.

Assisted-by: opencode:claude-opus-5-5
Document the pasta requirement, --publish / --publish-addr /
--host-port / --net-log, the per-user publish addresses and what they
do not guarantee, host.internal and the loopback-trust caveat, and the
descriptor relay.

Assisted-by: opencode:claude-opus-5-5
passt is in RHEL 9 but not RHEL 8. Without a working pasta the
sandbox now shares the host network as before, with a warning, instead
of refusing to start. --local-net chooses the host network explicitly
and silences the warning. The forwarding options still need pasta:
with --local-net they are an error, and without pasta they stop the
launch (exit 3). BWRAP_NET_MODE tells the sandbox which mode it is in,
and the context file describes both.

Assisted-by: opencode:claude-opus-5-5
Use bwrap's own term for sharing the host network namespace.

Assisted-by: opencode:claude-opus-5-5
Without a usable private network namespace the launchers now exit 3
naming the reason, instead of warning and sharing the host network.
Only --share-net shares it. BWRAP_NET_BACKEND (pasta or none) replaces
BWRAP_NET_MODE, ready for a second backend.

Assisted-by: opencode:claude-opus-5-5
RHEL 8 has slirp4netns but no passt. When pasta is not usable, the
launchers now use slirp4netns: util-linux unshare gives the launcher a
user namespace that owns a new network namespace, bwrap runs the
sandbox in it as the real user, and a background helper attaches
slirp4netns (--disable-host-loopback, 192.168.254.0/24, built-in DNS)
before releasing the sandbox through --block-fd. The launcher still
execs in the foreground; slirp4netns exits with bwrap (--sync-fd). On a
setup failure the helper kills the blocked sandbox before the tool
starts.

unshare, not bwrap --unshare-net, makes the namespaces: bwrap's devpts
setup nests the sandbox in a second user namespace, from which
slirp4netns would have no rights over the network namespace.

--publish works through the slirp4netns API socket; --host-port and
--net-log need pasta. --net-backend restricts the choice to one
backend.

Assisted-by: opencode:claude-opus-5-5
Document the selection order (pasta, slirp4netns, else stop unless
--share-net), --net-backend, what each backend supports, and how the
slirp4netns launch works.

Assisted-by: opencode:claude-opus-5-5
--serve now adds its port to --publish: the server binds 0.0.0.0 inside
the private namespace and is reached on the host at the user's publish
address 127.H1.H2.S (normally .1), with both backends. With --share-net
it binds 127.H1.H2.1 itself. The port is chosen free on that address,
the printed ssh -L command and pairing links name it, and
--serve-ssh-config forwards to it. net_publish_prefix exposes the
per-user prefix.

Assisted-by: opencode:claude-opus-5-5
The slirp4netns pipes and API socket were in a mktemp directory under
XDG_RUNTIME_DIR or TMPDIR, whose parent was not checked: another user
able to write it could swap the directory during the launch. They now
go in a per-launch session directory under $XDG_RUNTIME_DIR/bwrap-net,
else /tmp/bwrap-net-UID, used only when no other user can change its
parent and it is the user's own 0700 directory. TMPDIR is never used.

Both locations are blocked paths, so no sandbox of the user can mount
them and reach another sandbox's API socket.

Assisted-by: opencode:claude-opus-5-5
The slirp4netns helper kept polling after the launcher had gone (for
example when unshare failed) and then sent SIGKILL to the saved pid,
which by then could belong to another process. It now exits quietly as
soon as its parent is no longer the launcher, and before killing it
checks that the launcher is still its parent and that the sandbox pid
is the launcher's child.

Assisted-by: opencode:claude-opus-5-5
The publish address came from one ss snapshot, and the backend binds
it only after exec, so two launches at the same moment could pick the
same address and port. The choice is now made under flock in the
runtime directory, and each launch leaves a claim (launcher pid and
start time, host ports) that later launches respect; claims whose
process is gone are removed. --serve port claims also record the start
time, so a reused pid does not keep them alive.

Assisted-by: opencode:claude-opus-5-5
The header still said that without pasta the sandbox shares the host
network with a warning; the launch tries slirp4netns and otherwise
stops unless --share-net is given.

Assisted-by: opencode:claude-opus-5-5
pasta writes one log file, and the network log becomes an always-on
journal record in the next commit; a file the user names is easier to
edit, lose, or leave where the agent can write it.

Assisted-by: opencode:claude-opus-5-5
…m log

Every launch writes start and end records (tag bwrap-net, a session
id also given to the sandbox as BWRAP_SESSION) with systemd-cat, else
logger; with neither it warns and runs without the log. There is no
option to turn it off.

With pasta every flow is logged too: pasta runs with --debug and a
FIFO as --log-file (huge --log-size, as its rotation fails on a FIFO)
and as its stderr, and a supervisor outside the sandbox copies each
line to the log, holding the FIFO open so pasta never blocks. The
descriptor relay prints a marker and reopens the user's stderr, so
pasta's setup lines go to the log, and if pasta fails before the
marker its lines are shown on the terminal.

Assisted-by: opencode:claude-opus-5-5
tests/fake-net-run runs a launcher against fake pasta, slirp4netns,
unshare, systemd-cat and logger, where /dev/net/tun is missing (as in
a development sandbox). The fakes really create or join the namespaces
and forward TCP, so --publish, --serve and the network log can be
tested end to end; failure modes are read from $FAKE_NET_DIR.

Assisted-by: opencode:claude-opus-5-5
A --serve session could move to another session index when one of its
ports was busy on 127.H1.H2.1, and the --serve-ssh-config alias then
did not reach it. --serve now fixes the session's publish address to
.1: its port is chosen free and unclaimed there, a clash with another
--publish port stops the launch with the reason, and --publish-addr is
refused with --serve.

The fake pasta now handles comma-separated port lists and never starts
the command when its setup failed.

Assisted-by: opencode:claude-opus-5-5
pasta's -T takes no listening address ("Listening address not allowed
for -T"): it always listens on the namespace's loopback. So
--host-port PORT[:SANDBOXPORT] now passes -T SANDBOXPORT:PORT and the
agent reaches the host service at 127.0.0.1:SANDBOXPORT; host.internal,
127.0.42.1 and BWRAP_NET_HOST_ADDR are gone. That sandbox port is held
by pasta, so a --publish to it is refused, and the context file says
the agent cannot serve on it.

The fake pasta now rejects an address on -T/-U and listens on the
namespace's loopback, as the real one does; it had accepted the old
form, which is why the tests passed.

Assisted-by: opencode:claude-opus-5-5
pasta's debug log has about 34 lines per connection (state changes,
timers, ACK bookkeeping) plus a doubled setup dump, and only addresses.
lib/bwrap_netlog.py condense now turns it into one logfmt line when a
connection opens and one when it closes (with its duration), a fail
line for connections that never became active, and pasta's warnings
and errors; 5000 lines of a real session became 280.

To name destinations, bwrap_netlog.py dns runs as a DNS relay on
127.53.0.1:53 in the sandbox's network namespace, started by the
descriptor relay before bwrap (it waits until the relay listens, at
most 0.25 s) and dropping its capabilities after binding. It logs one
dns line per lookup and forwards to pasta's DNS address, which
resolv.conf lists second so lookups go on working if the relay is
down. The condenser adds name= to connections from the answers. Both
need Python 3.6+ (platform-python on RHEL 8); without it the log is
unfiltered as before.

Assisted-by: opencode:claude-opus-5-5
Unit tests (pixi run -e dev test) against part of a real pasta debug
log with documentation addresses, and a DNS relay against a local fake
upstream. The fake pasta now writes pasta's debug format for its
forwards and answers DNS on --dns-forward, so the launchers' log can
be checked end to end with tests/fake-net-run.

Assisted-by: opencode:claude-opus-5-5
On a real session the first IPv6 connections were logged without a
name: the client connected as soon as it had the answer, before the
relay had written its dns line. The relay now logs first. CNAMEs go
to cname= so answers= holds addresses only.

Assisted-by: opencode:claude-opus-5-5
…rk log

name= did not say what it was. Connection and dns lines now use
domain= (the domain the relay looked up), which is also what an
allow-list is written in.

Assisted-by: opencode:claude-opus-5-5
The upstream actionlint hook is language: golang: prek downloads a Go
toolchain and builds it in every fresh environment, which is where
sandboxed sessions' connections to go.dev, dl.google.com and
proxy.golang.org came from. conda-forge has actionlint 1.7.12, the
pinned version, so it now runs from the pixi dev environment like the
other linters.

Assisted-by: opencode:claude-opus-5-5
~/.cache/prek (or $PREK_HOME, passed through) is shared with the host
when it exists, like the npm, pixi and ccache caches. A sandbox no
longer sets up every hook environment again, and so no longer builds
gitleaks from source with a freshly downloaded Go toolchain (go.dev,
dl.google.com, proxy.golang.org) each session. As with the other
caches, the agent can change what is in it, and prek runs hooks from
it on the host.

Assisted-by: opencode:claude-opus-5-5

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

Human assessment is needed because directory validation can allow a symlink target containing GitHub credentials to be exposed through a read-write data mount.

1 open issue

Review effort: Balanced

Copilot AI balanced review requested due to automatic review settings November 14, 2026 02:03

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

Browser-accessible shell execution, agent-readable credentials, and optional host-network access require human security/operations review, DSSI approval, and managed-host validation.

1 open issue

Review effort: Balanced

Only the start and end records carried user=, so a dns, open, close or
fail line could be attributed only by looking up its session. Every
record now begins session=ID user=NAME, written by the condenser
(--user) and by the fallback supervisor used without Python.

The name is the account of the launcher process (id -un), not $USER,
which the caller can set, and characters that could split a logfmt
field are replaced. The start and end records use the same value.

Assisted-by: opencode:claude-sonnet-5-5
bwopenchamber sources this file for its OpenCode 2 mounts, environment and --serve port pool. Running the launcher, directly or through a symlink, is unchanged.

Assisted-by: opencode:claude-sonnet-5-5
OpenChamber is a web server that starts its own opencode serve and adds a terminal, git and worktree handling. All of that does work, so it all runs in one sandbox: the server, its OpenCode, the shells, git and every tool call. Only the browser is outside, behind an ssh tunnel.

The launcher sources bwopencode2 for the OpenCode 2 mounts, the database guard and the --serve port pool, so one ssh_config alias covers both. The port is published on the user's 127.H1.H2.1; the server listens on 0.0.0.0 inside the private network. A random UI password is set for each session.

Enterprise mode is forced on through the environment and /etc/openchamber/policy.json is not mounted, because a policy file would override it. OPENCODE_BINARY is pinned and no host OPENCODE_* or OPENCHAMBER_* variable reaches the server. OpenChamber's run/ directory, where it writes the UI password in plain text, is a tmpfs. No OpenChamber argument is passed through.

Assisted-by: opencode:claude-sonnet-5-5
SCRIPT_DIR came from $0, which is the running script. When bwopenchamber or a test sources bwopencode2 from another directory, the library was looked up next to that script, so it only worked by accident. BASH_SOURCE[0] is the same value when bwopencode2 runs directly or through a symlink.

Assisted-by: opencode:claude-sonnet-5-5
… limits

OpenChamber 2.7.0 signs sign-in cookies with a key it writes to jwt-secret in its data directory and reuses. A browser that signed in stayed signed in across restarts and after the password changed, and the agent could read the file. The launcher now sets a random OPENCODE_JWT_SECRET for each session; checked against the real server that the old cookie is refused after a restart and that no jwt-secret file is written.

The docs now say that the published 127.H1.H2.1 address is not access control (any local user can connect, the UI password keeps them out), which the first version of this launcher got wrong. They also record that the agent can read the UI and server passwords OpenChamber puts in its environment, checked in a running session, and that hiding them from the agent's shell would not help because /proc/PID/environ is readable by the same user. The agent is told in the sandbox context not to use them.

The checks made on a real launch move from not checked to checked. Two tests that could not fail are fixed: --dry-run prints variables named PASSWORD or SECRET as REDACTED, so they now check the names.

Assisted-by: opencode:claude-sonnet-5-5
Checked in a running session: the agent's shell has OPENCODE_JWT_SECRET set (printenv | grep -c prints 1). The per-session key stops a reusable key being stored, but an agent can still sign a cookie for the current session. Say so under the sign-in key and in the known limit.

Assisted-by: opencode:claude-sonnet-5-5
…ests independent of sessions

Copilot review of NSLS2#43 found two real problems, both checked.

Pinning. The launcher passed OPENCODE_BINARY and the docs called OpenCode pinned. Without a policy file OpenChamber 2.7.0 lets a saved opencodeBinary in settings.json override OPENCODE_BINARY, and settings.json is in the persistent data directory that the agent can write. A fake OpenCode named there ran. Only a policy file outranks the saved setting. The launcher now generates /etc/openchamber/policy.json and mounts it read-only from a descriptor (enterprise mode, the installed OpenCode pinned, allowNetworkAccess true for the private network, no extensions, built-in skills catalogues hidden). With it the installed OpenCode started and the fake never ran, and the server still listened on 0.0.0.0. OPENCHAMBER_ALLOW_NETWORK_ACCESS is dropped: the file governs, and with the file missing the server now fails closed instead of running unpinned.

Tests. The harness probed real sockets on the user's publish address, so a running bwopenchamber or bwopencode2 --serve session holding 24700 made every dry run pick 24701 and fail the checks that expect the first port. The probe is stubbed. Without the stub the same scenario gives 24701; with it 24700. New checks cover the generated policy and that a held port changes nothing.

Assisted-by: opencode:claude-sonnet-5-5
…etwork

The database guard section said the lock is a descriptor that the final exec passes on to bwrap. With a private network the launcher execs into pasta, which closes inherited descriptors, so nothing in the sandbox holds it. On a real launch fuser on the lock file showed only the network log's helper processes (a forked copy of the launcher shell and its condenser) holding it; it was held all session and gone after exit. So the guard works, but through inheritance by a helper started before the exec.

Record that, and the two cases the code implies but that were not tested: no log tool, so no helper, and a helper stopped mid-session. Prompted by a Copilot review summary that said pasta releases the lock: the claim is wrong for the normal launch and right in outline for those cases.

Assisted-by: opencode:claude-sonnet-5-5
…one on PATH

The launcher took the first node on PATH and stopped if it was older than 24.14. On RHEL 10 the default node is 22 and the nodejs24 package installs node-24 next to it, so a new enough Node was installed and the launcher still refused. Resolve it in this order: BWOPENCHAMBER_NODE if set (only that one; an error if too old), the node on PATH if it is new enough, then the newest node-24 or nodejs24 style binary in /usr/bin and /usr/local/bin that is. The error lists what was found with versions and what to do on RHEL 10, RHEL 8 and 9, and with BWOPENCHAMBER_NODE (for example a pixi Node).

Checked on the real host with no stubs: /usr/bin/node and /usr/bin/node-22 are both 22.23.2 and the message lists them; BWOPENCHAMBER_NODE set to a pixi Node 24.22.1 launches. The tests are hermetic by default (the versioned search points at an empty directory) and cover each rule, a binary that is not node and one that is not executable. RHEL 9 (runansible1) has 24.21.0 on PATH from the nodejs:24 module and RHEL 10 offers nodejs24 24.19.0 through Satellite; whether nodejs24 installs a binary called node-24 is inferred from node-22 being there, not checked.

Assisted-by: opencode:claude-sonnet-5-5
…x relay uses

oc2db_guard took the session lock on bash's first free descriptor (10 or 11). The launcher's descriptor relay hands the sandbox its files and data on descriptors 10 and up and opens them over whatever is there, closing the lock and so releasing it while the sandbox runs. bwopencode2 with --share-net (one data descriptor, the lock on 11) kept it by luck; bwopenchamber (policy 10, context 11, lock 11) lost it. Found in Copilot's review of NSLS2#43 and reproduced with the real relay shim and a real bwrap.

The lock is now on the first free descriptor from 200 up. A lock on 200 stayed held through the real relay and bwrap. With pasta the relay never sees it (pasta closes inherited descriptors first), so that path is unchanged. Two new checks open descriptors 10 to 29 through the real relay and assert the lock is above 199 and still held; both fail on the old code. 82 guard checks and the 95 bwopenchamber checks pass.

Assisted-by: opencode:claude-sonnet-5-5
…ne on PATH

The resolver treated the first candidate as the node on PATH and stopped there when it was new enough. With no node on PATH the first candidate is the first versioned binary (node-24 sorts before node-25), so it chose that one instead of the newest, which the docs promise. Track the node found on PATH explicitly and prefer only that. A test with a PATH that holds no node and node-24 and node-25 in the search directory fails on the old code (it picked node-24) and passes now; 97 checks pass. Found by Copilot's review of NSLS2#43.

Assisted-by: opencode:claude-sonnet-5-5
Print the ssh command, a one-time sign-in link and a google-chrome --app
command in one block once the server is up, and a new link when Enter is
pressed. --serve-open starts a browser on the host, outside the sandbox, with
a pairing link of its own.

Assisted-by: opencode:claude-sonnet-5-5
…ut a browser

With no browser installed and none registered, xdg-open on GNOME opens the
link in the text editor, which shows an Unauthorized error and uses up the
one-time link. Use xdg-open only when xdg-mime names a default web browser,
try firefox before it, and otherwise say that nothing will be opened.

Assisted-by: opencode:claude-sonnet-5-5
…ve link

The server redirects and sets the cookie only for Accept: text/html; curl or the GNOME text editor get a JSON token and use up the one-time code.

Assisted-by: opencode:claude-sonnet-5-5
…en browser

The helper inherited the database session lock descriptor and so did the browser it started, which then held the lock for as long as its window stayed open and made the next launch on the same data directory fail. Close every descriptor above 2 in the helper.

Also check the default browser for http, the scheme of the sign-in link, not https.

Assisted-by: opencode:claude-sonnet-5-5
--serve-open is now --serve plus a browser on the same host, so it needs no --serve next to it, and --serve alone still opens nothing. --serve-port also works with --serve-open.

Assisted-by: opencode:claude-sonnet-5-5
… blocked path

The data directory is bound whole and read-write, but it was only checked with _check_dir_safe, which accepts a symlink to a directory that contains ~/.config/gh. Check it in bind mode and mount the resolved path.

Assisted-by: opencode:claude-sonnet-5-5
For the host you are on, print one line when the browser opened, or only a fresh one-time link to open by hand with the password when it did not (no browser, no desktop session, the browser exited with an error within 8 seconds, or the server never answered). --serve prints the same block as before.

Assisted-by: opencode:claude-sonnet-5-5

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

The mount validation can expose service.json and stored environment values, and stopping the sandbox during pairing can still trigger a browser launch and a false success report.

1 open issue

Review effort: Balanced

actionlint became a local pixi hook, so its repo entry is gone and the grep exited 1, stopping every lint job in Collect tool versions under pipefail.

Assisted-by: opencode:claude-sonnet-5-5
Copilot AI balanced review requested due to automatic review settings November 16, 2026 00:44

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔵 Needs a closer look

The browser terminal requires human security/operations sign-off on documented credential exposure and --share-net tradeoffs, plus managed-host integration validation, while its sandbox/network behavior depends on unmerged PR #37.

No open issues 🎉

1 resolved since last review

Review effort: Balanced

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants