Repository navigation
v0.38.5
Bug fixes. The first start after upgrading takes longer than usual: see "Upgrading" below.
Upgrading
- Every agent harness on an existing volume is reinstalled once, at the first start after the upgrade, because each is now pinned and stamped (#646). That took 8 to 14 minutes on the amd64 hosts we measured.
- If the network fails during that start, the harnesses it couldn't reinstall are reported as not installed and are tried again at the next start. A harness isn't left running bytes that weren't checked.
- Later starts read each harness's stamp and run nothing.
HERMES_MCP_PINis no longer read. UseHR_HERMES_MCP_VERSIONwith an exact version; a start-time warning names the old variable if it is still set.HR_OPENCODE_VERSIONno longer acceptslatest.
What was wrong
- First-start installs followed upstream (#646, #623, thanks @eric-epsilla). opencode was fetched from
releases/latestwith no check, and Qwen Code and Pi were installed with no version. So a volume ran whatever upstream published on the day it was created.- Every first-start install is now pinned: opencode 1.18.23, Qwen Code 0.26.3, Pi 1.5.0, and exact versions for Hermes's and the other Python harnesses' top-level packages. Release assets are checked against digests committed in the image.
- A stamp replaces the version checks that used to run each installed harness at every start.
- A source check in CI fails on any unpinned install, so the list of exceptions can only shrink.
- An installer's processes now see only what fetching and building need: the path, the locale, proxy and CA settings, and your registry settings (
PIP_*,NPM_CONFIG_*,UV_*). They no longer see the instance's internal key, the console password, or provider keys passed with-e. - Installs still run as root, so this stops a package that reads or reports its environment, not one that sets out to read the instance's secrets. Running installs as their own user is the next step in #623.
- A connection couldn't send the extra header a gateway in front of the provider needs (#636, thanks @Mohith26). Examples are Cloudflare AI Gateway's
cf-aig-authorization, Helicone, Portkey, and OpenRouter's attribution headers.- A connection can now carry request headers, and every model call on it sends them after the key.
- A header value is held like the key. It is never handed to the agent, so a connection with headers is brokered even under
HR_SANDBOX_TRUST=owner. It is masked when read, and you type it again when the connection's address changes. - Names the gateway sets itself are refused, in any case. So are names that speak for a client or a hop (
via,forwarded,proxy-*,x-forwarded-*, Cloudflare's owncf-*markers) and AWS request signing (x-amz-*).
- A sandbox that stopped mid-task lost the task's steps (#634, #505, thanks @abhisheksharma001). With
HR_STEP_CHECKPOINT=1, the workspace is also saved after each finished tool step while a task runs.- It's off by default here, because a self-hosted workspace is on the volume and already outlives a restart. Turn it on if your workspaces live somewhere a restart loses.
- The settings are in the self-hosting guide's environment table.
- Setup builder (#649). On a deployment shared by several organizations, the setup builder now holds an organization's own connection address to the rule the model relay has used since 0.38.4. A single-operator install, the default, is unchanged.
Checked
On 0.38.5-rc.1, on an existing volume upgraded from 0.38.4:
- The upgrade reinstalled all 30 harnesses, each stamped by root, in 13 minutes 40 seconds with no warning. A second start came up in 13 seconds and installed nothing.
- On a fresh volume, the release gate installed all 30 on amd64 and arm64, and its second start installed and warned about nothing.
- Request headers: Pi tasks on a connection with headers sent them on every model call, after the key, and nothing the gateway uses internally went upstream.
- The values were nowhere on the volume outside the encrypted store, and not in the logs.
- A saved connection reads its values masked, an unchanged save keeps them, and the refused names are refused.
- Pi 1.5.0:
- A session started on Pi 0.84.2 resumed.
- New sessions on a chat completions and a Responses connection wrote a file and recalled it.
- An MCP tool call went through.
- A refused key reads as a failed task.
- On a deployment shared by several organizations, with real name lookups, names that resolve to a loopback, link-local or private address are refused before any socket opens, on the model relay and the setup builder. A public provider is reached at an address its name resolved to.
- The 0.38.3 and 0.38.4 checks pass again: the MCP probe's limits, a turn bound to its session, and Pi and Oh My Pi on a custom Responses connection.
- The release's tree is identical to the candidate's.
Full Changelog: v0.38.4...v0.38.5
Security advisory: GHSA-x35v-ppr8-m5vh. On an instance shared by several organizations, an organization's own connection address could reach internal networks: the model relay is fixed in 0.38.4, the setup builder in 0.38.5.