Meta Muse Spark 1.3 Contributor announces subagent invocation then stalls with no tool call

Bug: Meta Muse Spark 1.3 Contributor announces subagent invocation then stalls with no tool call

Model: opencode-go/muse-spark-1.3-contributor (also seen as opencode/muse-spark-1.3-contributor-free)
I am using Meta Muse Spark 1.3 Contributor.

OpenCode version: 1.18.31
OS: Linux
Install method: npm / binary (opencode-ai)

What happened

The model frequently responds that it is going to invoke / dispatch subagents (e.g. via the task tool / dispatching-parallel-agents workflow) and then stops all processing:

• It outputs text like "Using X to Y" / "I'll dispatch subagents to..." / announces parallel workers.
• Then no task tool call is ever made.
• The turn ends / stalls with no further tool calls, no file edits, no follow-up. Session appears hung waiting for work that never starts.

This happens repeatedly when the task would normally fan out to subagents (exploration, parallel recon, plan execution). It does not recover on its own — user has to re-prompt ("continue", "proceed with the tool call") and it often repeats the announcement without calling the tool.

Expected

If the model announces subagent dispatch, it should actually emit the task tool call(s) in the same turn, or if subagents are unavailable for the model/provider, it should do the work inline instead of stalling.

Actual

Announcement with no tool call, then stop. No error, no tool result, no recovery.

Repro

1. Use model opencode-go/muse-spark-1.3-contributor in opencode (default agent with task subagent tool available + skills that encourage subagents).
2. Ask for any multi-part task that triggers subagent fan-out, e.g. "explore codebase for X and Y in parallel" or any prompt where skills say to dispatch parallel agents.
3. Observe assistant message announcing subagent invocation, followed by end-of-turn with zero tool calls.

Provider config context

• opencode models lists the model as opencode-go/muse-spark-1.3-contributor.
• No custom model config in ~/.config/opencode/opencode.jsonc (default Zen / opencode-go provider path).
• Skills enabled: superpowers plugin + local skills path (dispatching-parallel-agents, etc.), which explicitly instruct task with subagent_type usage — model seems to narrate the skill instead of executing it.

Suspected cause

Model is narrating the tool-use plan (possibly from skill text like "dispatch parallel agents") as chat output instead of emitting the structured task tool call. Could be:

• tool-call formatting / stop-sequence issue specific to this model on opencode-go provider, or
• system-prompt/skill wording that this model interprets as narration rather than action.

A mitigation on the opencode side could be a nudge when a turn contains "dispatch/invoke subagent" language but zero tool calls, or a provider-specific fix for tool-calling with this model.

Logs

Happy to provide ~/.local/share/opencode/log/ + session storage on request — tell me which run/session ID format you want.

Start by reproducing the behavior with opencode-go/muse-spark-1.3-contributor and the task tool using the documented multi-part prompt. Inspect the provider/tool-calling path and the requested session logs to determine whether the model response is being emitted as text instead of a structured call; done means dispatch occurs or the turn recovers inline without stalling.