OpenCode and Pi are two of the coding engines that build cards on an OPVS desk. Before version 1.11.1 both read one context window and one output limit for the whole desk, measured for MiniMax-M3 only. Each build now uses the limits of the model it runs.
What's new in v1.11.1
- Breaking: a build with unknown limits stops and says why. When a build's model publishes no limits and is not the model the desk-wide values describe, the desk no longer guesses. The build stops before the engine starts and records the failure cause
harness_limits_unknown. Pin such a card to a model that publishes its limits. - Each build gets its own model's window and output limit. The credential a desk receives for a build now carries that model's context window and output limit. OpenCode and Pi compact the conversation and size each reply from those numbers, so a long build on a 262,144 token model no longer plans for a 1,048,576 token window and fails half-way.
- A reply always fits inside the window. Some models count the requested output inside the context window. The output limit is now capped at half the window less a 16,384 token margin, and compaction keeps that much room free, so a late turn is not refused for being too long.
- Proven on a live build. On 18 October an OpenCode build on Kimi K2.7 Code ran with a 262,144 token window, a 114,688 token output limit and 131,072 tokens held back for compaction. Those are Kimi's own limits, not MiniMax-M3's.
Desks get these changes with sidecar 0.2.130 or later.
Read every release note at https://opvs.ai/changelog.