PR Inbound :: Add OpenCode and Mistral Vibe backends
SwarmForge currently validates and launches a small fixed set of coding-agent CLIs. This makes it awkward to use OpenCode's provider/model selection or Mistral Vibe's authenticated CLI as role backends.
I have a focused fork branch that adds both backends:
opencode launches OpenCode with the assigned worktree, role arguments (including --model provider/model), and the generated role prompt.vibe launches Mistral Vibe programmatically in the assigned worktree with --trust --agent accept-edits --output streaming -p <role prompt>. Role arguments are passed through before -p, so a configuration can set explicit --max-turns, --max-tokens, and --max-price limits.- The Vibe launcher deliberately does not add
--auto-approve or --yolo.
- The branch also adds a bounded, opt-in smoke-only instruction profile and makes alternate project-local configurations selectable with
SWARMFORGE_CONFIG.
Validation so far was performed in a mountless aarch64 Lima guest. The Babashka suite passed there with 27 tests / 104 assertions. Direct bounded Vibe programmatic proof passed, including streaming output and an exact marker response.
Important limitation: a Vibe role launched in programmatic streaming mode terminates after its response. Its attempt to perform a SwarmForge handoff callback reaches Vibe's approval gate, but the callback is immediately denied in this mode; therefore I am not claiming an end-to-end Vibe handoff-delivery smoke has passed. The proposed PR documents and tests the safe launch contract only, not a benchmark or a completed Vibe receiver workflow.
The corresponding focused draft PR is #40.
If folks are totally against this please vote down. If you want me to finish up the PR then vote up. This is an age where no-one seems to bother with PRs any more which is yet more "dead internet": so I would do with the encouragement.
With respect to the PR at the moment it is for OpenCode Zen and OpenCode Go and Mistral.AI vibe as I am trying them out myself. I should (with sufficient encouragement) probably split it out so that Mistral.AI is a standalone PR as it is mostly likely most similar to what is already there. OpenCode is inherently more complex as it has multiple providers, a cheap $10 sub, plus pay-as-you-go, with different limits and sometimes different models. It is totally awesome as as with Zen they host open-weight models in US, EU and Singapore so excellent fun.
PR Inbound :: Add OpenCode and Mistral Vibe backends
SwarmForge currently validates and launches a small fixed set of coding-agent CLIs. This makes it awkward to use OpenCode's provider/model selection or Mistral Vibe's authenticated CLI as role backends.
I have a focused fork branch that adds both backends:
opencodelaunches OpenCode with the assigned worktree, role arguments (including--model provider/model), and the generated role prompt.vibelaunches Mistral Vibe programmatically in the assigned worktree with--trust --agent accept-edits --output streaming -p <role prompt>. Role arguments are passed through before-p, so a configuration can set explicit--max-turns,--max-tokens, and--max-pricelimits.--auto-approveor--yolo.SWARMFORGE_CONFIG.Validation so far was performed in a mountless aarch64 Lima guest. The Babashka suite passed there with 27 tests / 104 assertions. Direct bounded Vibe programmatic proof passed, including streaming output and an exact marker response.
Important limitation: a Vibe role launched in programmatic streaming mode terminates after its response. Its attempt to perform a SwarmForge handoff callback reaches Vibe's approval gate, but the callback is immediately denied in this mode; therefore I am not claiming an end-to-end Vibe handoff-delivery smoke has passed. The proposed PR documents and tests the safe launch contract only, not a benchmark or a completed Vibe receiver workflow.
The corresponding focused draft PR is #40.
If folks are totally against this please vote down. If you want me to finish up the PR then vote up. This is an age where no-one seems to bother with PRs any more which is yet more "dead internet": so I would do with the encouragement.
With respect to the PR at the moment it is for OpenCode Zen and OpenCode Go and Mistral.AI vibe as I am trying them out myself. I should (with sufficient encouragement) probably split it out so that Mistral.AI is a standalone PR as it is mostly likely most similar to what is already there. OpenCode is inherently more complex as it has multiple providers, a cheap $10 sub, plus pay-as-you-go, with different limits and sometimes different models. It is totally awesome as as with Zen they host open-weight models in US, EU and Singapore so excellent fun.