Repository navigation
v0.31.0
New
- An Azure connection can sign in with Microsoft Entra instead of an API key. For an organization that issues no API keys for its Azure resources. On the Integrations page, add an Azure OpenAI connection and set "Sign in with" to "Microsoft Entra", then enter the endpoint, the directory (tenant) ID, the application (client) ID and a client secret of an application registered in your own directory with the "Cognitive Services OpenAI User" role on the resource. Tasks sign in as that application. The agent never holds the secret or the token: the service signs in for it, on every model call, and signs in again before the token expires, so a task that runs for hours keeps working. When the sign-in is refused, the task stops at once and says so in Microsoft's own words (a wrong directory, application or secret, or an application with no role on the resource), and it does not move on to another connection. The self-hosting guide has the
azcommands.
Changes
- A model call that sends nothing for 3 minutes is ended (it was 10). When a provider accepts a request and then goes silent, the task is now told after 180 seconds. Most models send events while they think, so this only shortens the wait on a call that is truly stuck. A few do not: in our measurement, gpt-5.5 and gemini-3.1-pro through TokenRouter's Chat Completions route, and claude-opus-5.5 through Vercel, sent nothing for 201 to 211 seconds before answering a hard problem at their highest thinking level, and such a call now ends as "the provider did not answer". If you run those models at those levels, set
HR_RELAY_UPSTREAM_TIMEOUT_S=600to keep the old wait. This change was taken back in 0.31.1, where the wait is 10 minutes again: a wider measurement showed the silence is not limited to those routes. Use 0.31.1 or later. - The Vercel deploy tool says when a preview became production. Vercel makes a project's first deployment a production one whatever is asked. An agent that asked for a preview is now told so in the tool's answer, with what to do next, instead of concluding its task is blocked.
Fixes
- A task started the moment an environment finished building is no longer refused. A build reported itself ready an instant before its version became the active one, and a task started in between failed with "the environment has no built version". A build now makes its version active first. If a build fails at its very last step, the environment keeps the version it had.
Checked
On this release's candidates, in the default owner trust, with a test application in our own directory:
- Pi, goose, Hermes and Codex each completed a task on an Azure OpenAI resource through an Entra connection, and the service signed in once for the four.
- While an agent was running, none of the container's processes and none of 5,995 workspace files held the client secret or an access token.
- One task ran across a renewal: it was started 2 minutes before the sign-in was due again and made a model call every 85 seconds for 9 minutes. The service signed in again at the third call, the last two calls came after the first token had expired, and the task completed.
- An application with no role on the resource, and a directory Microsoft does not know, each failed the task with Microsoft's reason (in 2 seconds on Pi).
- An Azure connection that uses an API key works as before.
- The connection form at 1440, 1024, 768 and 390 pixels wide.
- An environment with a pip, an npm and an apt package built in 9 seconds, and a task started the instant it reported ready completed. The environments check passed on Pi, Codex and Claude Code, 9 of 9 tasks.
- Claude Code tasks complete, the thinking check passes on Codex and Pi, and all plugin checks pass on Claude Code, Codex and Pi.
Not checked: a completed task on an Azure AI Foundry resource on this build (the resource accepted the application's sign-in; the hosted service completed a task there with the same code), a sovereign cloud, and a client secret that is rotated while a task runs.
How long each model stayed silent, one hard problem at its highest thinking level (seconds; the longest stretch with nothing sent):
| Model | TokenRouter | Vercel |
|---|---|---|
| gpt-5.5, Chat Completions | 201 | 11 |
| gpt-5.5, Responses | 30 | 61 |
| gpt-6.1-sol | 19 | 15 |
| gemini-3.1-pro-preview | 205 | 4.5 |
| claude-opus-5 | 35 | |
| claude-opus-5.5 | 211 | |
| grok-4.6 | 13 | 12 |
| deepseek-v4-pro | 2 | |
| kimi-k3 | 0.7 | |
| qwen3.7-max | 2 |
One call (gpt-6-sol through TokenRouter) never answered in 20 minutes.
docker run -d -p 3000:3000 -v hr-data:/data harnessrouter/harnessrouter:0.31.0
Full Changelog: v0.30.2...v0.31.0