[Feature]: Orchestration — support opencode agents (e.g. Muse Spark) as workers, not only Claude/Codex/Cursor

[jondmarien]

Problem or use case

Teams standardized on opencode + Muse Spark (e.g. a session running on muse-spark-1.3-contributor via opencode) cannot use Orca supervised orchestration (Run/Task/Dispatch, worker_done settlement) with their own models. worker-start only launches Claude/Codex/Cursor agents:

• orca-ide orchestration worker-start --help (verified on orca-ide 1.4.199, Linux, 2026-09-10):
  • --agent <agent> = "Launch a known TUI agent" — there is no --agent value for opencode agents.
  • --model <id> = "Provider model id for a new agent launch", with the note "--model supports Claude, Codex, and Cursor opaque provider model ids" — no --model namespace for opencode models, so the coordinator cannot spawn same-model workers and had to fall back to its own subagents.
• orca-ide account list shows managed accounts exist only for Claude/Codex ("Managed Claude accounts (1)", "No managed Codex accounts") — no opencode account/managed-auth surface.
• The agent-context --json schema (234 commands) confirms the same three-agent world.

What was attempted: orca-ide orchestration worker-start --task <id> --agent ... / --model ... with an opencode model id — rejected / unsupported, since only Claude/Codex/Cursor provider model ids are accepted.

Expected: a coordinator running inside opencode (Muse Spark) can spawn supervised workers running the same opencode model, with the same Run/Task/Dispatch lifecycle and worker_done settlement as Claude/Codex/Cursor workers.

Actual: no --agent opencode launch path and no provider-ID passthrough for non-Claude/Codex/Cursor models, so opencode-based teams are locked out of supervised orchestration.

Proposed solution

1. Add an --agent opencode launch path for orchestration worker-start (launch the opencode TUI agent in the worker terminal, same as the existing known agents).
2. Accept generic/opaque provider model IDs via --model for opencode models (e.g. Muse Spark IDs), or document the passthrough if it already exists.
3. Document the valid --agent values in worker-start --help so callers don't have to guess which agents are supported.
4. If managed auth applies, surface opencode accounts in account list (or document that opencode uses its own auth and needs no managed account).

Alternatives or additional context

• Workaround today: coordinator falls back to its own subagents (no supervision, no Run/Task/Dispatch settlement) — loses exactly the durability/audit properties orchestration exists for.
• Minimal acceptable slice: (a) document valid --agent values in help, plus (b) either an opencode launch path or generic provider-ID passthrough for --model.
• Related: README lists opencode among supported agents/topics side-by-side, so orchestration worker support would close a gap between the marketed agent matrix and the orchestration agent matrix.
• Environment: orca-ide 1.4.199, Linux, verified 2026-09-10.

[nwparker]

Closed by #22383: orca orchestration worker-start --agent muse --model <id> --effort <level> now launches Muse Spark workers with the requested model and reasoning effort, under the full Run/Task/Dispatch lifecycle and worker_done settlement. So a coordinator can spawn same-model Muse workers directly.

Not covered: --model for opencode workers. The opencode 2.x terminal UI rejects --model (only opencode run accepts -m), so selecting an opencode model at launch would have to go through OPENCODE_CONFIG_CONTENT, which worker launch preferences can't set yet. --agent opencode without --model works and uses opencode's configured model. If you need per-worker opencode model selection, please open a follow-up.