Summary

When the orchestrator runs a multi-agent workflow, every sub-agent's full output (tool calls, thinking blocks, intermediate transcript, final answer) gets concatenated into the parent's context. With 5–10 sub-agents, the parent turn balloons fast — slow API requests, mid-turn compaction churn, and a UI that looks frozen because each round-trip is huge. This is structural, not lock contention; it happens even if #510 fixes the freeze.

Pairs with #510 — both make multi-agent workflows feel broken, for different reasons.

Two structural fixes

1. Summarize sub-agent results at the boundary

When a sub-agent finishes, the parent currently receives the raw agent_result payload — full transcript, every tool call, every thinking block. Replace with a structured summary:

• Final answer / deliverable (verbatim).
• A short narrative of what the sub-agent did, not the raw transcript.
• File paths touched + what changed (one line each).
• A full_log_ref token pointing to the spillover file (OPENCODE: Tool-output spillover file with retention #422) so the parent can request the full log on demand.

This is the same pattern compaction.rs uses for the parent's own history — apply it at the sub-agent boundary.

2. Don't keep agent-card progress in the model's context

The Agents sidebar / agent_card widget is for the user. It shouldn't flow into the model's context as a running history. Right now (verify) each agent state change ends up in the orchestrator's transcript window. With 5 agents transitioning through pending → running → step N/100 → completed, that's hundreds of low-information lines.

Keep progress events in the UI/event stream only. The model gets one event per terminal transition with the summary from (1).

Likely areas

• crates/tui/src/tools/subagent/mod.rs — sub-agent result payload constructor.
• Wherever the result tool returns its payload.
• runtime_threads.rs / runtime_api.rs — split the UI-event stream from the model-context stream so progress doesn't bleed in.
• compaction.rs — possibly extend with a sub-agent-specific compaction template (matches the spirit of OPENCODE: Compaction template - structured Markdown #429).

Related

• v0.8.8 tui: full UI freeze during multi-agent fan-out (likely SubAgentManager mutex contention) #510 (lock contention), v0.8.8 agents: completed sub-agents still count toward the running cap; default of 5 too low #509 (cap counting) — same release; all three needed for multi-agent to work.
• OPENCODE: Tool-output spillover file with retention #422 (spillover file) — provides the full_log_ref target.