Repository navigation
A security patch over 1.18.0, from the findings the reviews of 1.18.0 left open. Codex works again under
docker-compose.host-subscriptions.yml, where every call had failed since 1.15.0, and its readiness now checks that it
can make its call directory. The controller no longer follows what a provider CLI leaves in its own home: a FIFO or a
link at Codex's config.toml, a link in a path the config names, a call directory moved aside. The CLI sandbox now
checks where the controller keeps its stores, and its own code, as it checks its secrets, follows the links on the
way to each path, and quotes the paths it logs. A token in BROWSER_WS_ENDPOINT's query, userinfo or fragment no
longer reaches the log. browser.get_local_storage reads in the controller's isolated world and bounds its answer;
the TOTP autofill presses no Enter on a page that left its host while the code was typed; a share link minted without
a lifetime lives SHARE_TOKEN_TTL_MINUTES; and a tunnel released twice no longer frees a newer tunnel's port. The
release workflow and audit refuse a CI count they cannot read, a lock that pins nothing, an advisory they cannot read
and more secret shapes. The controller's dependencies are updated (FastAPI 0.162.4, pydantic 2.14.3).
Read the upgrade notes first: a copy of the host-subscription override kept outside this repository needs the new
tmpfs entry; with CLI_SANDBOX_REQUIRED=true, a store outside /data, a host run from a checkout or virtualenv a CLI
may write, or a writable entry or CLI_HOME reached through such a place now stops provider CLIs; a Codex
config.toml that is a link or larger than 1 MiB is not read; a share link minted without a lifetime follows
SHARE_TOKEN_TTL_MINUTES where that is not 60; and every browser.get_local_storage read answers with
truncated.
The controller image changes, and the browser-node image in how it redacts an endpoint it prints; the LangChain
package changes in its error redaction; the client and MCP bridge packages change only in version. Advisories:
GHSA-9qq7-x8h7-87fv and GHSA-363m-96wq-fxmg, fixed in this release.
Upgrade notes
-
A copy of
docker-compose.host-subscriptions.ymlkept outside this repository needs the new tmpfs entry. The
override now mounts an empty tmpfs atCLI_HOME(see Fixed); the override in this repository needs nothing. A copy
made before 1.18.1 still leaves Codex unable to make its call directory where the controller does not run as root,
as by default, and the controller now says so: Codex's
readiness is not ready, with the reason, and withAPP_ENV=productionandOPENAI_AUTH_MODE=clithe controller
refuses to start, Claude and Gemini with it, where before it started and every Codex call failed. Add the entry to
the copy'svolumes, beside the four read-only sign-ins:- type: tmpfs target: ${CLI_HOST_HOME:-/home/botuser} tmpfs: size: 67108864 # 64 MiB
Never set
CLI_HOST_HOMEto/rootwith it: every user in the container could then write root's home there. -
A Codex
config.tomlthat is a link, or larger than 1 MiB, is no longer read. The Codex path-compatibility
shim readsCLI_HOME/.codex/config.tomlonly as a regular file of at most 1 MiB that is itself at that name (see
Security), and logs a warning for anything else there; the.codexlinks such a file names are not made. Put the
file itself inCLI_HOME/.codex, at most 1 MiB, and restart the controller: a file it refused is not looked at
again before the next start. On Windows, which cannot open a file without following a link, a link there is still
read. -
A
.codexpath inconfig.tomlthat passes through a link inCLI_HOMEis no longer linked. Where a
directory on such a path insideCLI_HOME, before its.codex, is a link (CLI_HOME/projectslinked elsewhere),
the shim now makes nothing through it and logs a warning that names the path (see Security); a.codexalready at
the path's end, a link or not, is left as it is. Make that directory a plain one, or make the.codexlink
yourself. On Windows, which cannot open a directory this way, the shim works by path as before. -
With
CLI_SANDBOX_REQUIRED=true, a store kept outside/datanow stops provider CLIs. A Linux host run that
keepsAPPROVAL_ROOT,STATE_DB_PATHor another of the controller's stores (see Security) outside/data, in a
directory a CLI may write, or holding one (aCLI_SANDBOX_WRITABLEentry inside it), ran its CLIs with that store
open to them and said nothing. It now gets the sandbox's warning, unless an earlier one comes first (see
Security), and withCLI_SANDBOX_REQUIRED=trueno provider CLI runs (Provider request failed). Keep the stores
under/data, as the image does, with no writable entry inside one. -
With
CLI_SANDBOX_REQUIRED=true, a host run from a checkout or a venv a CLI may write now stops provider CLIs.
Where the checkout, the virtualenv or the interpreter lies in/tmp,/var/tmp,CLI_HOMEor an entry of
CLI_SANDBOX_WRITABLE, or holds such an entry (see Security), the CLIs ran, and a CLI could change the code the
controller runs. Now the sandbox's warning names the place, and withCLI_SANDBOX_REQUIRED=trueno provider CLI
runs. Move the checkout and its virtualenv out of those directories, and keep the entries out of them. -
A
CLI_SANDBOX_WRITABLEentry or aCLI_HOMEwhose way to its target passes through/tmp, or another
directory a CLI may write, is no longer given. Such an entry, a link to a link in/tmpsay, is now passed over
with a warning that says where it resolves (one that leads into a directory the CLI is given anyway,/tmpitself,
needs no warning: that directory's rule covers it), and such aCLI_HOMEis said to be refused, as one that is a
link in/tmpwas already; withCLI_SANDBOX_REQUIRED=truethe refused home stops provider CLIs (see Security).
Set the entry or the home to its real path, or keep the links on its way out of the directories a CLI may write. -
A share link minted without a lifetime now lives
SHARE_TOKEN_TTL_MINUTES, where that is set to other than
60. Such a link lived 60 minutes whatever the setting said (see Fixed); it now lives as long as the setting says,
and one set above 1440 is read as 1440, with a warning at startup. A client that wants 60 minutes there asks for
ttl_minutes: 60.ttl_minutesmay now benull, which means the same as leaving it out, and its default in the
request's and the tool's schema isnull, not 60. -
browser.get_local_storagereads what the browser holds, and cuts an answer past 1,000,000 characters. A page's
own replacement of its storage functions (Storage.prototype.getItem,window.localStorage) no longer changes what
the read returns, and an answer is cut at 1,000,000 characters (see Security). Every answer that reads the store
now carriestruncated,truewhere something was cut or left out. Read a store that large one key at a time.
Security
- A FIFO at Codex's
config.tomlno longer holds the controller. The Codex path-compatibility shim reads
CLI_HOME/.codex/config.tomlat the first Codex call after each start, on the controller's event loop, and it
opened whatever was at that name and waited for it. A CLI may make a FIFO there, in its own home, and the read
then waited for a writer: the controller answered nothing meanwhile (on Linux, nothing ran for the 5 seconds a
probe waited). The shim now reads the file as the controller reads a call'sdecision.json: only a regular file of
at most 1 MiB that is itself there, opened without waiting. A FIFO, a link, a directory or a larger file is not
read, and a warning says which it was, once. On Windows a link there is still followed. - The Codex path-compatibility shim no longer makes directories through a link in
CLI_HOME. For a path under
CLI_HOMEthatconfig.tomlnames (CLI_HOME/x/deep/.codex/...), the shim makes any missing directories and a
.codexlink to the real one. It made them by path, and the controller is not under the CLI sandbox: a CLI writes
that config and can put links in its home, so withCLI_HOME/xlinked to/data/auththe controller made
directories of the CLI's naming in/data/auth, where the sandbox keeps the CLI out. The shim now opens each part
from the one before it and makes the next one there, never through a link: a part that is a link when the shim
comes to it, or that a CLI puts back as one before the shim opens it, stops that path with a warning naming it (the
directories made before it stay), and one swapped after the shim opened it leads nowhere, as the shim goes on in
the directory it opened. Paths through plain directories the controller may read are linked as before; on Windows,
which cannot open a directory this way, the shim works by path as before. - A Codex call's directory moved aside by a CLI no longer leads the controller elsewhere. The controller makes
each Codex call's working directory inCLI_HOME, makes the call's schema file in it, reads the call's
decision.jsonfrom it and removes it after, all outside the CLI sandbox. It did each of these by path, so a CLI
working in the home meanwhile (another call's, at the same time) could move the directory aside and leave a link at
its name: the controller then made the schema file in the place the link led to, took adecision.jsonthere for
the call's decision, and failed the call when it came to remove the directory. 1.18.0 named this as a limit. The
controller now makes the directory from a descriptor ofCLI_HOME, opens it at once without following a link,
and keeps that descriptor: the
schema file is made, anddecision.jsonread, in the directory it made, wherever that has been moved, and the
directory is removed without following a link. Where a link has been left at its name the directory stays, a
warning says so, and the call's outcome stands; so does one that holds a directory the CLI made read-only, which
1.18.0 opened up by path to remove. On Windows, which cannot open a directory this way, all of it is done by path
as before. - The CLI sandbox now checks where the controller keeps its stores, as it checks its secrets. It warned where
API_TOKEN_FILE,AUTH_ROOTor another secret's setting named a place outside/data, or one a CLI may write,
and said nothing of the stores: withAPPROVAL_ROOTorSTATE_DB_PATHkept outside/data, a sandboxed CLI could
read the approvals' digest key or the controller's state, underCLI_SANDBOX_REQUIRED=truetoo, and the log had
no line about it.APPROVAL_ROOT,AUDIT_ROOT,STATE_DB_PATH,SESSION_STORE_ROOT,JOB_STORE_ROOT,
MEMORY_ROOT,HARNESS_ROOT,MESH_PEERS_PATH,MCP_SESSION_STORE_PATH,CRON_STORE_PATH,
SHARE_REVOCATION_STORE_PATH,ISOLATED_TUNNEL_KNOWN_HOSTS_PATH,ISOLATED_TUNNEL_INFO_ROOT,
REMOTE_ACCESS_INFO_PATH,ARTIFACT_ROOTandUPLOAD_ROOTare now checked the same way: the warning names each
one that is left open, andCLI_SANDBOX_REQUIRED=truerefuses to run the CLI. The sandbox names one condition at a
time, the first it finds: a refused home, then the controller's code, then the secrets and stores, so a store left
open is named once the two before it are put right. Kept in their default places under
/data, as in the image, nothing changes. - The CLI sandbox now reports a host run whose own code a CLI could change, and refuses it under
CLI_SANDBOX_REQUIRED=true. The launcher that starts each provider CLI under the sandbox is a file of the
controller's package, and the controller's interpreter runs it before any ruleset. On a Linux host run whose
checkout or virtualenv lies in a directory a CLI may write (/tmp,/var/tmp,CLI_HOME, the CLI'sTMPDIR, an
entry ofCLI_SANDBOX_WRITABLE), a sandboxed CLI could add to the launcher, and the next call ran what it added
outside the sandbox, with no warning, underCLI_SANDBOX_REQUIRED=truetoo; the same held for the rest of the
package, the interpreter and the installed packages at the controller's next start. The sandbox's check of what
the controller reads at its start now takes in its own code: its package with the launcher, its interpreter as it
was started and once resolved,sys.prefix,sys.base_prefixand where its packages are installed. Where a CLI
may write one of them the warning names it, andCLI_SANDBOX_REQUIRED=truerefuses to run the CLI; without that
setting the CLI runs after the warning, as after the sandbox's other warnings, so set it on a host run. The image
keeps all of these out of what a CLI may write. - A
CLI_SANDBOX_WRITABLEentry that leads through a place a CLI may write is no longer given where it leads.
The sandbox gives each entry as its resolved path. It looked at the entry as written, at where it resolves and at
each leading part of it resolved, but not at the links on the way: an operator's link outside every directory a
CLI may write, to a link in/tmp(/opt/xto/tmp/sub/inner), was given at wherever/tmp/sub/inner
pointed, and a CLI could point it at any directory the controller's user may write, the controller's own code
among them, and be given that to write at the next call. Each place Linux looks up on the way now counts too, for
the entries, the CLI home and the secrets' and stores' paths alike: such an entry is passed over with a warning
that says where it resolves, and a secret reached that way is reported as open. - A directory name with a line break in it no longer starts lines of its own in the controller's log. The CLI
sandbox's warnings name the directories they are about, as written and, where it can be read, where each
resolves. A CLI may make a
directory with a line break in its name in its own home and point a link it can swap at it (one an operator's
CLI_SANDBOX_WRITABLEentry resolves through, say); the next call's warning then carried the name as it was, and
what followed the line break read as a log line of its own. The warnings now quote each path a setting or a
lookup gave them, its line breaks escaped. - A token in
BROWSER_WS_ENDPOINT's query, userinfo or fragment no longer reaches the controller's log. The
controller names the endpoint it connects to in three places: the log line at every connect, the connect failure
message (which names a configuredBROWSER_WS_ENDPOINT, and which readiness, deep health and the startup log
write to the log) and the line that logs Playwright's own connect error. Each replaced only the endpoint's path
with<redacted>, so a hosted browser's endpoint with its token anywhere else (wss://host?token=...,
wss://user:token@host/..., a#fragment), or with its scheme in capitals (WSS://), was logged whole, token
and all. Every Playwright endpoint the
controller prints (the browser-node's, orBROWSER_WS_ENDPOINT) now keeps only its scheme, host and port, with
/<redacted>where it had anything more; one whose host and port cannot be read plainly is printed as its scheme
alone (ws://<redacted>). So is an endpoint in the browser sandbox details of/readiness, which went through
the same redaction; the browser-node's log and sandbox report and the CI sandbox scripts redact the same way. - A session's tunnel released a second time no longer frees a port another session's tunnel holds. With
ISOLATED_TUNNEL_ENABLED=true, each session's takeover tunnel reserves a remote port on the bastion, and its
release frees the port. A tunnel released again, as two closes of one session made together could do, freed its
port again, whichever tunnel held it by then: the next session's tunnel was then started on a port a newer session's
tunnel still held. A release now acts only on the tunnel the broker holds for that session: one released
already, or given up while it started, is left alone, and one whose teardown failed is torn down again at its next
release or at shutdown. - The TOTP autofill no longer presses Enter on a page that left the named host while the code was typed. After
each step that waits on the page (the search for the field, its focus click, its clearing), the autofill checks
again that the page is on a host intotp_hostsand not refused. After the typing it did not, so a page that a
key handler sent to another host as the code was typed had Enter pressed into it, which submits whatever form has
the focus there. It checks there too now, just before Enter, and such a page gets no Enter; one that leaves in
the moment between that check and the key is not caught. browser.get_local_storageno longer returns what the page chooses, nor an answer of any size. It read the
storage in the page's own JavaScript world, so a page that replacedStorage.prototype.getItemanswered the read
itself: 20 million characters in a probe, every one returned to the agent, far more than a page's storage can hold.
The read now runs in the controller's isolated world, asbrowser.get_htmlwithtext_onlyand
browser.find_elementsdo, where the page's own JavaScript does not answer for its storage. A store can hold
megabytes by itself all the same, and a key as long as a value, so the answer is bounded too: its keys and values,
counted as the answer writes them (a quote, a backslash or a line break takes two characters there, any other
control character six), with eight characters more for each entry, come to at most 1,000,000 characters, the most
any tool'smax_charsaccepts, and an answer cut there saystruncated: true. The tool is in thestoragepack,
which the default profile leaves out.
Fixed
- Codex works again under
docker-compose.host-subscriptions.yml, broken since 1.15.0. 1.15.0 moved the
directory each Codex call works in from the system temp dir intoCLI_HOME. The override setsCLI_HOMEto the
host-style home (/home/botuserby default) and mounts the host's.codex,.claude,.claude.jsonand
.geminiread-only inside it, but nothing made the home itself: the container runtime made it, as the parent of
those mounts, owned by root with mode 0755, and the controller, which does not run as root there, could not make
a directory in it. Every Codex call failed (Agent step failed, and in the logcould not make a working directory under the CLI home) while the provider list showed Codex ready. The override now mounts an empty 64 MiB tmpfs at
that home (mode 1777,noexec, emptied at every start), and the four sign-ins stay read-only inside it. A CI step
runs the override with/bin/trueas Codex and fails where the call directory cannot be made. - Codex's readiness, and a production start, now check that a directory can be made in
CLI_HOME. Readiness
asked only whetherCLI_HOMEexisted and held.codex, so a home the controller could not write showed ready.
Codex is now not ready there, with the reason (codex cannot make its call directory under CLI_HOME /home/botuser (Permission denied), naming the setting's path and not the one the error carried), and with
APP_ENV=productionandOPENAI_AUTH_MODE=clithe controller refuses to start. The check makes one directory and
removes it at once; where a CLI working in the home has filled or removed it meanwhile, the home still counts as
writable and the log has a warning that names no path (one it replaced with another empty directory is removed
without one). The Claude and Gemini CLIs make no such directory and are
not checked. - The release audit's secret scan finds more of OpenAI's and Slack's key shapes, and AWS secret access keys. It
found an OpenAI key writtensk-orsk-proj-and then the key, but not OpenAI's service account, admin and user
keys (sk-svcacct-,sk-admin-,sk-None-), nor Slack'sxoxc-andxoxe-tokens, nor an AWS secret access
key, so a key of these kinds left in a file the audit reads passed it. It finds them now. An AWS secret access key
has no prefix of its own, so the scan finds one where it follows its setting's name (aws_secret_access_keyor
AWS_SECRET_ACCESS_KEY, then=or:); AWS's documented example secret key is let through only as the PII
scrubber's tests write it (aws_secret_access_key=and the key), as its example key ID already was. - The release workflow no longer publishes to PyPI on a count of green CI runs it could not read. Its gate counts
the successfulci.ymlpush runs on main for the tagged commit and stops at none. A count that came back as no
number (null, or nothing) failed the gate's comparison as a test, which theifaround it read as false, so the
gate passed and the packages went on to PyPI with no green run read. Such a count now stops the release, as it
already stopped the image workflow. - The release audit's Python dependency audit fails on a lock that pins nothing for the image. It hands
pip-audit the pins ofcontroller/constraints.txtthat hold for the image, and pip-audit given none audits
nothing and passes: an emptied lock, or one whose every line is for another platform, passed the audit. The audit
now stops there and says so. - The release audit checks that the MCP metapackage depends on the client by the client's name. It took any
Requires-Distline that began withauto-browser-client, so a wheel that depended on another package whose name
begins so (auto-browser-client-typo) passed, as did one whose README held such a line. It now reads the wheel's
dependencies from its metadata's headers and compares each one's name as pip does, soauto_browser_clientstill
counts andauto-browser-client-typodoes not. - The release audit's advisory check blocks on an advisory it cannot read. It blocked only on an advisory whose
state wastriage, so one with no state, with its state under another key, or with a state GitHub might add read
as clear; an entry that was not an object ended the check with a traceback, and a--from-jsonfile holding an
empty object, not a list, passed. Each of these now blocks the release and says to check the advisories by hand. The
states it knows are GitHub's five:triage,draft,published,closedandwithdrawn. - The LangChain integration's errors read a wrapped controller URL as the MCP bridges do. The integration keeps a
copy of the bridges' redaction, as it depends on httpx alone. The bridges read the URL as urllib does, without a
<URL:...>or<...>wrapper around it; the copy kept the wrapper, so for such a base URL it left the query's and
the fragment's secrets in a text the bridges redact. httpx refuses such a URL, with an error that holds no secret,
so nothing was shown; the copy now reads it as the bridges do. The test that holds the copy to the bridges needs the
client, which the integration's CI job did not install, so it ran in no CI job and missed this; the job now
installs the client. - A share link minted without a lifetime lives as long as
SHARE_TOKEN_TTL_MINUTESsays. A mint that named no
ttl_minutes(POST /sessions/{id}/sharewith no body or without it,browser.share_sessionwithout it) made a
link for 60 minutes whatever the setting said: the route and the tool each put 60 in place of the lifetime not
given, so a deployment's own default, a shorter one included, never applied. They now leave it to the setting (60
unless set). A setting longer than a caller may ask for, 1440 minutes, is read as 1440, with a warning at startup:
a link minted without a lifetime lives no longer than one minted with one.
Changed
- Controller dependencies: FastAPI 0.162.4 (
>=0.162.4,<0.163), pydantic 2.14.3, pydantic-core 2.50.4,
typing-inspection 0.4.5 and click 8.5.1, which 1.16.0 held back as under 72 hours old, and annotated-doc 0.0.6,
filelock 4.2.10, markdown-it-py 4.3.0 and py-serializable 2.2.0 (#273);controller/constraints.txtresolved again
under a cutoff of 2026-11-12. Dev: ruff 0.17.4. - The image workflow's
docker/login-action4.8.0 (#268).
Tests
Thirteen guards that held, but that no test would have missed had they been removed, now each have a test that fails
without them: the cut of a cookie's name and destination in a write's record; the kill of a provider CLI whose output
passes its cap; the approval a fast-profile click still asks for where the button's fuller name is too long to read;
a secret field's text hidden in three more approval answers; the selector quoted in the log line of a CDP element read
that fails; the TOTP autofill's check after its focus click; the dashboard's origin check before it sends its token
for a replay's screenshot; the store of revoked share links the controller is built with; the bound on the read of a
Codex call's decision.json; the sandbox flags on each Codex path's codex exec; the approval store's refusal of an
id that leads out of APPROVAL_ROOT; the loss a pre-request gate reports after its browser failed to close; and a
governed tool call's approval counting only while the call holds it. Nothing they cover changes.