Result Synthesis

Result Synthesis

Close the loop on an orchestration.plan: gather each agent's result, resolve conflicts, speak one answer, and make sure the holograms they spawned are arranged comfortably (not piled on top of each other).

Inputs

• The plan_id and the per-agent results (each agent's data + the holo.spawn

objects it produced, addressed by object_id).

• The final orchestration.agent_status{state:"done"|"failed"} for every node.

Steps

1. Collect & order results in the order most useful to the user (usually:

what they asked first → supporting detail). Use each tool's data.speech suggestion as raw material, not verbatim.

1. Reconcile conflicts / dedupe. If two agents answered overlapping

questions, keep the higher-confidence / more specific one.

1. Compose one spoken reply. Stream long answers as multiple

agent.speech{final:false} and end with a single agent.speech{final:true}. Keep it conversational; don't read JSON aloud.

1. Coordinate the layout. Prefer delegating to the stage-agent

(compose-workspace) so it decides anchoring/arrangement. For a quick gather you may emit holo.layout yourself via arrange_holograms.

1. Mark the turn done — emit agent.thinking{stage:"done"} after the final

speech.

Output examples

Spoken synthesis (after perception + research + productivity finished):

{ "text": "That's your coffee mug. Tokyo is 18° and cloudy, and your 5-minute timer is running.",
  "final": true, "emotion": "neutral" }

Final layout via the stage compositor (arrange_holograms → holo.layout, protocol §5.10):

{ "arrangement": "arc", "anchor": "head", "spacing": 0.25,
  "objects": ["O_vision_annotation", "O_weather_orb", "O_timer"] }

Partial failure

If a node reports state:"failed", synthesize around it honestly — report what succeeded and name what didn't, without dropping the whole turn:

{ "text": "Tokyo is 18° and cloudy and your timer's set. I couldn't get a clear
  look at the desk object though — want me to try again?", "final": true }

Edge cases

• Nothing succeeded → one apologetic agent.speech{final:true} + an optional

notify toast; do not spawn data widgets.

• Too many holograms → ask stage-agent to run declutter-space before the

final arrangement.

• Barge-in (client.barge_in, §5.14) arrived mid-synthesis → stop streaming

speech, keep already-spawned holograms, end the turn.

• Streaming order → never emit final:true until every kept result is voiced.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sumitaich1998
• Source: sumitaich1998/jarvisvr
• License: MIT

Install and usage instructions live in the source repository linked above.