The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
Rust agent-orchestrationai-agentsai-assistantdesktopllmlocal-firstmcppersonal-aiprivacyrustsecond-braintauri
test: update agent harness e2e scripts to use spawn_subagent
Steven Enamakel8bbb0763e6c8
diff --git a/app/src/features/conversations/tools/toolSpecs.ts b/app/src/features/conversations/tools/toolSpecs.tsindex 5bc8afb8da..46c7d9bffa 100644--- a/app/src/features/conversations/tools/toolSpecs.ts+++ b/app/src/features/conversations/tools/toolSpecs.ts@@ -63,7 +63,6 @@ import {MailXIcon,MapPinIcon,MessageCircleQuestionIcon,- MessageSquareReplyIcon,MousePointerClickIcon,NetworkIcon,NewspaperIcon,diff --git a/tests/agent_harness_e2e.rs b/tests/agent_harness_e2e.rsindex 195e0872ba..d28e7bfc70 100644--- a/tests/agent_harness_e2e.rs+++ b/tests/agent_harness_e2e.rs@@ -1110,13 +1110,13 @@ async fn subagent_delegation_happy_path_inner() {// The orchestrator selects task_manager_agent through spawn_subagent, and// ask_user_clarification is in that agent's named tools. A child that needs a// detail pauses and hands the parent a `[SUBAGENT_AWAITING_USER]` envelope as-// the `manage_tasks` tool result (#4291). That envelope is an ordinary tool-// result to the orchestrator — `manage_tasks` is not in its `early_exit_tools`+// the spawn_subagent result (#4291). That envelope is an ordinary tool result+// to the orchestrator — spawn_subagent is not in its `early_exit_tools`// — so the orchestrator relays the question in its OWN next reply, and turn 1// ends on that text.//// Actual LLM request ordering (4 upstream calls total):-// request[0] = orchestrator turn 1 → manage_tasks (blocking) tool call+// request[0] = orchestrator turn 1 → spawn_subagent tool call// request[1] = task_manager_agent → ask_user_clarification pauses the child// request[2] = orchestrator, with the awaiting-user envelope in context →// relays the question as text; turn 1 ends (chat_done)@@ -1137,14 +1137,17 @@ async fn delegated_clarification_flow_inner() {let _lock = env_lock();reset_script(vec![// ── turn 1 ──- // request[0]: Orchestrator delegates to task_manager_agent via manage_tasks.+ // request[0]: Orchestrator selects the task manager.tool_call_completion(- "manage_tasks",- json!({ "prompt": "Add my GitHub issues as a task source", "blocking": true }),+ "spawn_subagent",+ json!({+ "agent_id": "task_manager_agent",+ "prompt": "Add my GitHub issues as a task source"+ }),),// request[1]: task_manager_agent asks for the missing detail. This pauses// the child; the question comes back to the orchestrator inside the- // `[SUBAGENT_AWAITING_USER]` envelope as the manage_tasks result.+ // `[SUBAGENT_AWAITING_USER]` envelope as the spawn_subagent result.tool_call_completion("ask_user_clarification",json!({ "question": "WHICH_VERSION_CANARY?" }),@@ -3369,8 +3372,11 @@ async fn agent_installs_a_registry_skill_then_runs_it_inner() {reset_script(vec![// Orchestrator hands the install to `skill_setup` (a direct hand-off, #6302).tool_call_completion(- "setup_skills",- json!({ "prompt": "Install the harness registry skill", "blocking": true }),+ "spawn_subagent",+ json!({+ "agent_id": "skill_setup",+ "prompt": "Install the harness registry skill"+ }),),// skill_setup finds the entry, then installs it (the gate parks here).tool_call_completion(@@ -3621,7 +3627,10 @@ async fn assert_packed_hand_off_reaches_specialist(packed_tool_call_completion(pack,hand_off,- json!({ "prompt": format!("Handle this through {hand_off}"), "blocking": true }),+ json!({+ "agent_id": hand_off,+ "prompt": format!("Handle this through {hand_off}")+ }),),text_completion("The specialist finished the task."),text_completion("The specialist handled it."),