Repository navigation
Replies: 2 comments 4 replies
|
Thanks for starting this, @smota. It's the kind of thread other people running OpenRig can learn from. A few places that may help while you plan the next round:
If you're open to it, sharing your own setup here or on openrig.dev/rigs would help others too. |
|
Sharing the current RepoStew topology in response to the question here. This is a customized Pi deployment on the native-Windows experimental branch, not a claim about supported Windows operation or a performance benchmark. We run six independent coordinator/worker groups: 16, 8, 8, 8, 8 and 6 workers respectively, plus one coordinator per group (60 allocated seats). Groups cover new-repository work, forge feedback, existing completed-candidate publication, and followed-repository new issues. Each group retains its own workers; a project has one writer even when two source streams mention it. The coordinator stays small: screen inputs, dispatch a complete bounded unit, consume the result, then continue/refill. The worker owns implementation, relevant validation and authorized delivery. We use six result decisions: delivered; no change needed; feasible work remains with the same worker; concrete blocker; unknown outcome requiring reconciliation; repeated execution failure needing recovery. A result message, queue closure and whole-domain completion are distinct. Context is split into an individually assembled fixed seat contract and a variable task/facts block. The fixed contract contains the relevant role/domain methods, avoiding repeated full-skill-tree reads. The variable block carries the current source/cursor, exact batch reference when applicable, ownership and unresolved effects. New repository assignments can then be short without losing their completion contract. Fresh coordinator context gets reconciled facts rather than the entire predecessor's reasoning. OpenRig remains the owner of queue, wake and lifecycle. Our Pi extension supplies application guidance and a read-only view of authorized worker capacity; it does not introduce another scheduler or task database. Global idle count is advisory, not a reservation or permission to borrow another group's worker. Explicit Stop and unknown writes survive handovers and model changes. What the latest check proves: 60 allocated sessions were running and their persisted model/thinking selections matched the requested configuration; all six expansion controls were consumed. It does not prove 60 simultaneous productive jobs, sustained throughput, or that every campaign is complete. The main practical lesson so far is to keep task authority, runtime activation and actual delivery separately observable. AI-assisted field report, posted at the operator's request. No private task transcripts or instance dumps included. |
Uh oh!
There was an error while loading. Please reload this page.
Hi everyone,
I’ve been running OpenRig for about a week now and really enjoying it. As I prep my next round of experiments, I’d love to see how the community is dialing in their setups.
If you're open to sharing:
What rig topology are you running?
How have you structured your culture?
Always keen to learn from different approaches and get some inspiration for the next iteration.
Cheers,
Sam
All reactions