diff --git a/src/openhuman/agent/harness/session/builder/builder_tests_part_01_tests.rs b/src/openhuman/agent/harness/session/builder/builder_tests_part_01_tests.rs
index 31231e8249..a9ae92c319 100644
--- a/src/openhuman/agent/harness/session/builder/builder_tests_part_01_tests.rs
+++ b/src/openhuman/agent/harness/session/builder/builder_tests_part_01_tests.rs
@@ -525,6 +525,26 @@ async fn build_session_agent_uses_profile_memory_instead_of_root_memory() {
 /// #6040 — the memory-access instruction is about the memory tools, not the
 /// learning subsystem, so it must be in the prompt with `learning.enabled`
 /// off (the default) whenever a retrieval tool is registered and visible.
+///
+/// Passes the definition explicitly via [`builtin_def`] rather than letting the
+/// factory resolve `"orchestrator"` from the registry, and that is load-bearing
+/// rather than ceremony.
+///
+/// The section is gated on `memory_recall` being **registered and visible after
+/// tool filtering** (`any_tool_offered`), and the visible set comes from the
+/// resolved definition's tool scope. With `None` here the factory reads
+/// `AgentDefinitionRegistry`'s `static GLOBAL: OnceLock<…>`
+/// (`harness/definition_part_02.rs:24`) — first-write-wins and never reset — so
+/// the test was asserting against whichever definition set some *other* test in
+/// the binary had installed first. That is exactly the hazard `builtin_def`
+/// was written for: it loads fresh from the bundled TOML, "entirely independent
+/// of the global registry singleton".
+///
+/// It is why this passed run alone and failed inside the full
+/// `openhuman::agent` run (`ci-lite` scopes the Rust lane per changed domain,
+/// so the whole scope only runs when a PR touches `agent/`), and why the
+/// sibling write-side test in `builder_tests_part_03_tests.rs` never flaked —
+/// it already supplied `builtin_def("orchestrator")`.
 #[tokio::test]
 async fn memory_access_instruction_is_present_with_learning_disabled() {
     use crate::openhuman::agent::context::prompt::LearnedContextData;
@@ -535,11 +555,20 @@ async fn memory_access_instruction_is_present_with_learning_disabled() {
     let mut config = test_config(&tmp);
     config.learning.enabled = false;
 
-    let agent = Agent::build_session_agent_inner(&config, "orchestrator", None, None, false, None)
-        .expect("build session agent");
+    let orchestrator = builtin_def("orchestrator");
+    let agent = Agent::build_session_agent_inner(
+        &config,
+        "orchestrator",
+        Some(&orchestrator),
+        None,
+        false,
+        None,
+    )
+    .expect("build session agent");
     let prompt = agent
         .build_system_prompt(LearnedContextData::default())
         .expect("build_system_prompt");
+
     assert!(
         prompt.contains(MEMORY_ACCESS_INSTRUCTION.trim()),
         "the memory-access section must not be gated on learning.enabled"
diff --git a/src/openhuman/agent/harness/subagent_runner/ops/graph_part_02.rs b/src/openhuman/agent/harness/subagent_runner/ops/graph_part_02.rs
index 0d01d0e379..3ccf7d3fcc 100644
--- a/src/openhuman/agent/harness/subagent_runner/ops/graph_part_02.rs
+++ b/src/openhuman/agent/harness/subagent_runner/ops/graph_part_02.rs
@@ -76,6 +76,16 @@ fn append_worker_message(
         "agent_id": agent_id,
         "task_id": task_id,
         "mode": "typed",
+        // #5934: a mirror row that is not the sub-agent speaking is a tool
+        // result, and the renderer maps every non-`"agent"` sender to
+        // `role: 'user'` (`assistantUiMessages.ts`) — so without this flag a
+        // tool's raw output is painted, in an openable worker thread, as
+        // something the human typed. `hidden` keeps the row in the log for the
+        // process rail and the process-source view; it only stops being chat.
+        // Genuine worker-thread user turns are written by
+        // `worker_thread::{create_worker_thread, append_worker_user_message}`,
+        // not here, so they are unaffected. Overridable by `metadata` below.
+        "hidden": sender != "agent",
     });
     if let (Some(base), Some(extra_fields)) = (extra.as_object_mut(), metadata.as_object()) {
         for (k, v) in extra_fields {
diff --git a/src/openhuman/agent/harness/subagent_runner/ops/graph_tests.rs b/src/openhuman/agent/harness/subagent_runner/ops/graph_tests.rs
index cbcadb68a3..f7ecc431b3 100644
--- a/src/openhuman/agent/harness/subagent_runner/ops/graph_tests.rs
+++ b/src/openhuman/agent/harness/subagent_runner/ops/graph_tests.rs
@@ -636,3 +636,109 @@ async fn an_allowlist_that_readmits_a_spawn_tool_is_refused_loudly() {
         "the readmitted spawn tool warns exactly once: {warnings:?}"
     );
 }
+
+/// #5934 (item 1): a mirrored sub-agent **tool result** must not be painted as
+/// something the human typed.
+///
+/// Both worker-thread mirrors write tool output with `sender: "user"`, and the
+/// renderer maps every non-`"agent"` sender to `role: 'user'`
+/// (`app/src/providers/assistantUiMessages.ts:536`) while keying visibility
+/// only on `extraMetadata.hidden` (`:664`, and the same flag in
+/// `ChatThreadView.tsx:312` / `timeline/selectors.ts:50`). Worker threads are
+/// openable chats — `create_worker_thread` stamps `labels: ["tasks"]` and
+/// `threadFilter.ts` lists that tab — so an unflagged mirror row shows the
+/// tool's raw output in a right-aligned user bubble.
+///
+/// The invariant: every non-`"agent"` row these mirrors write is `hidden`. The
+/// record stays in the log for the process rail; it just stops being chat.
+#[test]
+fn mirrored_tool_results_are_hidden_from_the_worker_thread_chat() {
+    use crate::openhuman::memory::conversations::{self as store, CreateConversationThread};
+
+    let dir = std::env::temp_dir().join(format!("wt-5934-{}", uuid::Uuid::new_v4()));
+    std::fs::create_dir_all(&dir).unwrap();
+    let now = chrono::Utc::now().to_rfc3339();
+    for id in ["worker-typed", "worker-recovered"] {
+        store::ensure_thread(
+            dir.clone(),
+            CreateConversationThread {
+                id: id.to_string(),
+                title: "task".to_string(),
+                created_at: now.clone(),
+                parent_thread_id: Some("parent-1".to_string()),
+                labels: Some(vec!["tasks".to_string()]),
+                personality_id: None,
+            },
+        )
+        .unwrap();
+    }
+
+    const RAW: &str = "{\"events\":[{\"title\":\"raw tool output the human never typed\"}]}";
+
+    // The typed path (`ConversationMessage::ToolResults`), used on a normal run.
+    mirror_worker_thread(
+        &dir,
+        "worker-typed",
+        "researcher",
+        "task-1",
+        &[
+            ConversationMessage::AssistantToolCalls {
+                text: Some("checking the calendar".to_string()),
+                tool_calls: vec![crate::openhuman::inference::provider::ToolCall {
+                    id: "call-1".to_string(),
+                    name: "list_events".to_string(),
+                    arguments: "{}".to_string(),
+                    extra_content: None,
+                }],
+                reasoning_content: None,
+                extra_metadata: None,
+            },
+            ConversationMessage::ToolResults(vec![
+                crate::openhuman::agent::messages::ToolResultMessage {
+                    tool_call_id: "call-1".to_string(),
+                    content: RAW.to_string(),
+                },
+            ]),
+        ],
+        Some("Here is your week."),
+    );
+
+    // The error-recovery path (`role: "tool"`), used when a run fails mid-turn.
+    mirror_worker_thread_from_history(
+        &dir,
+        "worker-recovered",
+        "researcher",
+        "task-1",
+        &[
+            ChatMessage::assistant("checking the calendar"),
+            ChatMessage::tool(RAW),
+        ],
+        Some("[subagent run failed before completion]"),
+    );
+
+    // Collected, not asserted per row: both mirrors must be reported, so a
+    // failure names every path that is still painting a tool result as chat.
+    let mut painted_as_user_chat: Vec<String> = Vec::new();
+    for thread_id in ["worker-typed", "worker-recovered"] {
+        let rows = store::get_messages(dir.clone(), thread_id).unwrap();
+        assert!(
+            rows.iter().any(|r| r.content == RAW),
+            "{thread_id}: the mirror wrote no tool-result row to assert on"
+        );
+        for row in rows.iter().filter(|r| r.sender != "agent") {
+            if row.extra_metadata.get("hidden").and_then(|v| v.as_bool()) != Some(true) {
+                painted_as_user_chat.push(format!(
+                    "{thread_id}: sender={:?} not hidden: {}",
+                    row.sender, row.content
+                ));
+            }
+        }
+    }
+    assert!(
+        painted_as_user_chat.is_empty(),
+        "mirrored tool results render in a user chat bubble:\n{}",
+        painted_as_user_chat.join("\n")
+    );
+
+    let _ = std::fs::remove_dir_all(&dir);
+}