Repository navigation
v0.31.1
Changes
- A model call may again stay silent for 10 minutes before it is ended (0.31.0 made it 3). If you upgraded to 0.31.0, upgrade again. We shortened this wait in 0.31.0 after measuring three routes where a model thinks without sending anything. A wider measurement the same day showed it is not three routes: OpenAI's Chat Completions API sends nothing at all while a reasoning model thinks, whether you reach it directly, on Azure or through a reseller, and most agents speak that API. On such a route a call that is thinking and a call that is stuck look the same, so a short wait ends both. Set
HR_RELAY_UPSTREAM_TIMEOUT_Sif you would rather end a stuck call sooner.
Checked
The same hard problem as a real task (Pi, gemini-3.1-pro-preview through TokenRouter, thinking level high):
| Wait | Result |
|---|---|
| 180 seconds, as in 0.31.0 | The model call was cut four times, the agent retried each time, and the task failed after 737 seconds with "the provider did not answer". |
| 600 seconds, the default again | Five runs of five completed, after 28, 125, 179, 212 and 494 seconds. |
How long each route stayed silent on that problem at the model's highest thinking level (seconds, Chat Completions, each call watched for 430 seconds):
| Model | TokenRouter | Vercel | OpenRouter | The vendor itself |
|---|---|---|---|---|
| gpt-5.5 | over 430 | 11 | 61 | OpenAI over 430, Azure over 430 |
| gemini-3.1-pro-preview | 187, 234 | 4.5 | 4.8 | Google 141 |
| claude-opus-5.5 | 211 | 10.6 | not measured |
On OpenAI's Responses API the same gpt-5.5 sent something at least every 30 seconds on all three routes; Codex, Hermes, OpenCode and Kilo Code speak that API.
On this release's candidate, Pi, Codex and Claude Code tasks complete and all plugin checks pass on the three.
docker run -d -p 3000:3000 -v hr-data:/data harnessrouter/harnessrouter:0.31.1
Full Changelog: v0.31.0...v0.31.1