mirror of
https://github.com/multica-ai/multica.git
synced 2026-08-07 11:14:28 +02:00
61898e808e
handleModelList resolved the discovery executable from cfg.Agents[provider] only, while runTask launches a custom runtime profile's own command_name. On a host running both, the model picker advertised the built-in binary's catalog for a profile runtime that executes a different (often differently versioned) binary, so an unsupported model was only rejected once the task actually ran. On a custom-only host the list failed outright with 'no agent configured for provider' even though the profile binary existed. Mirror runTask's resolution order: profile spec first, built-in provider entry (with MUL-4486 self-heal) second, and only fail when neither exists. Also key the 60s discovery memo by (provider, executable path) for every dynamic-discovery provider. codex/opencode/deveco already did; the rest shared one entry per provider, so a built-in CLI and a same-family custom profile on the same host could serve each other's catalog. Co-authored-by: multica-agent <github@multica.ai>