spawn_parallel_agents: opt out of the 120s per-tool timeout so the fan-out isn't truncated (#4686)

src/openhuman/agent_orchestration/tools/spawn_parallel_agents.rs

@@ -14,7 +14,9 @@ use crate::openhuman::agent_orchestration::spawn_parallel_graph::{
     SpawnParallelGraphOutcome, SpawnParallelTaskValidationError,
 };
 use crate::openhuman::tinyagents::current_run_cancellation;
-use crate::openhuman::tools::traits::{PermissionLevel, Tool, ToolCallOptions, ToolResult};
+use crate::openhuman::tools::traits::{
+    PermissionLevel, Tool, ToolCallOptions, ToolResult, ToolTimeout,
+};
 use async_trait::async_trait;
 use serde_json::json;
 use tinyagents::harness::tool::ToolExecutionContext;
@@ -95,6 +97,20 @@ impl Tool for SpawnParallelAgentsTool {
         PermissionLevel::Execute
     }
 
+    /// Run **without** the global per-tool wall-clock deadline. This is a
+    /// fan-out primitive: it spawns N independent sub-agents and awaits them
+    /// concurrently via `spawn_parallel_graph`'s bounded `map_reduce`. Under the
+    /// default `Inherit` policy the whole fan-out is hard-killed at the
+    /// single-tool timeout (120s by default) — so a group of long workers is
+    /// truncated at one worker's budget instead of running to ~the slowest, and
+    /// each worker's research is cut short. The fan-out is already bounded
+    /// internally — by `max_concurrency`, the run cancellation token, and each
+    /// sub-agent's own iteration/turn caps — so it governs its own lifetime,
+    /// like the long-running scripting tools (`shell`, `node_exec`).
+    fn timeout_policy(&self, _args: &serde_json::Value) -> ToolTimeout {
+        ToolTimeout::Unbounded
+    }
+
     async fn execute(&self, args: serde_json::Value) -> anyhow::Result<ToolResult> {
         self.execute_with_context(args, ToolCallOptions::default(), None)
             .await

src/openhuman/agent_orchestration/tools/spawn_parallel_agents_tests.rs

@@ -16,6 +16,7 @@ use crate::openhuman::inference::provider::{
     ChatRequest, ChatResponse, ConversationMessage, Provider, ToolCall,
 };
 use crate::openhuman::memory::{Memory, MemoryCategory, MemoryEntry, NamespaceSummary, RecallOpts};
+use crate::openhuman::tools::traits::ToolTimeout;
 use crate::openhuman::tools::{PermissionLevel, Tool, ToolResult};
 use async_trait::async_trait;
 use parking_lot::Mutex;
@@ -899,3 +900,15 @@ async fn agent_turn_runs_long_parallel_subagent_flow_with_many_nested_tool_calls
         "each subagent should run three tool calls plus a final completion iteration"
     );
 }
+
+#[test]
+fn spawn_parallel_agents_opts_out_of_the_global_tool_timeout() {
+    // A fan-out of N long sub-agents must not be hard-killed at the single-tool
+    // wall-clock default (120s): that truncates every worker and bounds the whole
+    // group at one worker's budget. The tool governs its own lifetime via its
+    // internal max_concurrency, cancellation token, and per-sub-agent caps.
+    assert_eq!(
+        SpawnParallelAgentsTool::new().timeout_policy(&json!({})),
+        ToolTimeout::Unbounded,
+    );
+}