Introducing Voyage, a new long-running mode for Diffex, for ambitious coding tasks: porting a framework, designing a product, or building one from zero to one. The bet is that frequent compaction with active recall matters more than a large context window alone.
Give Voyage a destination
Voyage works best with an outcome the agent can verify.
Start with /plan, then give the plan to Voyage with /goal in the Diffex CLI.
Benchmark setup and cost
On 15 CPU-only SWE-Marathon v1.1 tasks, eight trials each, Voyage with GPT-6 Luna reached a 22.5% full-pass rate, 4.2 percentage points above the same model in Codex. Its 79.1% mean partial score at $1.54 per trial nearly matched Kimi K3’s 79.3% at about $850. Partial credit measures progress, not a completed task.

Voyage context management
Built on Pocket Context, Voyage lets the agent choose what to compact and when. Original evidence stays archived; a recall subagent searches it and returns a cited answer when details are needed.
Try Voyage
Read the full announcement on the Diffex blog for installation, benchmark details, and the implementation.