diff --git a/crates/op-orchestrator/src/concurrent.rs b/crates/op-orchestrator/src/concurrent.rs
index 986a4910d..c7bfc5dd4 100644
--- a/crates/op-orchestrator/src/concurrent.rs
+++ b/crates/op-orchestrator/src/concurrent.rs
@@ -1,16 +1,34 @@
-//! S3b-2 Task A1: concurrency decision + screen grouping.
+//! S3b-2 Tasks A1 + B1: concurrency decision + screen grouping + worker.
 //!
-//! Port of `orchestrator.ts:780-810`.
+//! Port of `orchestrator.ts:780-810` (Task A1) and
+//! `orchestrator-sub-agent.ts:248-312` (Task B1 worker body).
+//!
+//! ## Task A1
 //! - `DesignRequest.concurrency` (added in `types.rs`)
 //! - `clamp_concurrency` — defensive [1, 6] clamp
 //! - `group_subtasks_by_screen` — group subtasks by screen; faithful to TS L785-801
 //! - `effective_concurrency` — concurrency decision; faithful to TS L803-810
 //!   (minus append-mode gate, which is S3b-4)
 //!
+//! ## Task B1
+//! - `BufferDocSink` — a per-worker buffering `DocSink` implementation that
+//!   collects applied `EditorCommand`s into a `Vec` without touching the real
+//!   document.  This is safe because `run_subtask` does NOT call `sink.state()`
+//!   at all — it only calls `sink.apply()`.  (Verified in `subagent.rs`.)
+//!   A snapshot of a pre-concurrent `EditorState` is still carried so that
+//!   `state()` can return something valid if any future caller does read it.
+//! - `run_screen_group_worker` — async fn that runs one screen group's subtasks
+//!   in order using the 2-attempt concurrent retry ladder (vs the sequential
+//!   path's 3-attempt tier-gated ladder in `run.rs`).
+//!
 //! Callers land in later S3b-2 tasks; scaffolding symbols are allowed to be unused.
 #![allow(dead_code)]
 
-use crate::plan::Subtask;
+use crate::plan::{OrchestratorPlan, Subtask};
+use crate::retry::is_non_retryable;
+use crate::subagent::run_subtask;
+use crate::types::{AbortFlag, DesignRequest, DocSink, LlmClient, Progress, SubtaskOutcome};
+use op_editor_core::{EditorCommand, EditorState};
 
 /// A group of subtask indices that share the same screen.
 /// `screen` is the screen name; `indices` are into the plan's `subtasks` slice.
@@ -89,6 +107,171 @@ pub(crate) fn effective_concurrency(concurrency: u32, screen_group_count: usize)
     }
 }
 
+// ── Task B1: BufferDocSink ────────────────────────────────────────────────────
+
+/// A per-worker buffering [`DocSink`] that collects every applied
+/// [`EditorCommand`] into an in-memory `Vec` without touching the real document.
+///
+/// **Design choice (spec §5 implementer note):** `run_subtask` (in `subagent.rs`)
+/// never calls `sink.state()` — it only calls `sink.apply()`.  Therefore the
+/// simplest correct buffer sink is a plain command collector; no document
+/// re-execution is needed.  We still carry a pre-concurrent `EditorState`
+/// snapshot so that any unexpected future `state()` call gets a valid (if
+/// stale) answer rather than a panic or empty document.  The snapshot should
+/// be taken just before the concurrent phase starts (after the N-root scaffold
+/// is applied to the real sink) — that is Task B2's responsibility.
+///
+/// After `join_all` the orchestrator replays `commands` into the real
+/// `DocSink` in subtask-plan-index order (Task B2: serialized replay).
+pub(crate) struct BufferDocSink {
+    /// Snapshot of `EditorState` taken before the concurrent phase.
+    /// Returned by `state()` unchanged — workers do not read it.
+    snapshot: EditorState,
+    /// All `EditorCommand`s collected via `apply()` calls.
+    pub commands: Vec<EditorCommand>,
+    /// Tracks undo-batch nesting depth (for parity with `DocSink` contract).
+    pub batch_depth: i32,
+}
+
+impl BufferDocSink {
+    /// Create a new buffer sink from a pre-concurrent state snapshot.
+    pub(crate) fn new(snapshot: EditorState) -> Self {
+        Self {
+            snapshot,
+            commands: Vec::new(),
+            batch_depth: 0,
+        }
+    }
+}
+
+impl DocSink for BufferDocSink {
+    fn state(&self) -> &EditorState {
+        &self.snapshot
+    }
+
+    /// Buffer the command for later replay; always returns `true`.
+    ///
+    /// We return `true` unconditionally — the real document is not touched
+    /// here; the per-worker result is validated after replay (Task B2).
+    fn apply(&mut self, cmd: EditorCommand) -> bool {
+        self.commands.push(cmd);
+        true
+    }
+
+    fn begin_undo_batch(&mut self) {
+        self.batch_depth += 1;
+    }
+
+    fn end_undo_batch(&mut self) {
+        self.batch_depth -= 1;
+    }
+}
+
+// ── Task B1: concurrent screen-group worker fn ────────────────────────────────
+
+/// Runs all subtasks for one screen group using the **2-attempt concurrent
+/// retry ladder** (vs the sequential path's 3-attempt tier-gated ladder).
+///
+/// Port of `orchestrator-sub-agent.ts:248-312`.
+///
+/// ## Retry ladder
+/// | Attempt | `reduced_complexity` | `minimal_skills` |
+/// |---------|----------------------|------------------|
+/// | 1       | `false`              | `false`          |
+/// | 2       | `true`               | `true`           |
+///
+/// Skips the sequential path's tier-gated middle attempt.
+///
+/// ## Retryable-failure gate (identical to sequential path)
+/// `error.is_some() && node_count == 0 && !abort.is_set() && !is_non_retryable(&err)`
+///
+/// Partial results (`node_count > 0`) are **never** retried.
+///
+/// ## Progress
+/// `progress_fn` is called with `SubtaskStarted` before each subtask and
+/// `SubtaskDone` / `SubtaskFailed` after.  In Task B2 this will be wired to
+/// an `mpsc::UnboundedSender` so multiple workers can fan-in; for B1 a plain
+/// `&mut dyn FnMut(Progress)` suffices.
+///
+/// ## Return value
+/// A `Vec<SubtaskOutcome>` with one entry per subtask index in `group.indices`,
+/// in the same order as `group.indices` (i.e. plan order within this screen).
+pub(crate) async fn run_screen_group_worker(
+    group: &ScreenGroup,
+    plan: &OrchestratorPlan,
+    request: &DesignRequest,
+    llm: &dyn LlmClient,
+    abort: &AbortFlag,
+    sink: &mut BufferDocSink,
+    progress_fn: &mut dyn FnMut(Progress),
+) -> Vec<SubtaskOutcome> {
+    let mut outcomes: Vec<SubtaskOutcome> = Vec::with_capacity(group.indices.len());
+
+    for &idx in &group.indices {
+        // Abort check at the top of every iteration (spec §4.4).
+        if abort.is_set() {
+            break;
+        }
+
+        let subtask = &plan.subtasks[idx];
+
+        progress_fn(Progress::SubtaskStarted {
+            id: subtask.id.clone(),
+            label: subtask.label.clone(),
+        });
+
+        // --- Attempt 1: full complexity ---
+        let outcome1 = run_subtask(subtask, plan, request, llm, sink, abort, false, false).await;
+
+        // Evaluate the non-retryable predicate once from attempt-1's error
+        // (matches TS `isNonRetryable` computed before the retry chain).
+        let non_retryable = outcome1
+            .error
+            .as_deref()
+            .map(is_non_retryable)
+            .unwrap_or(false);
+
+        // Retryable iff: error + zero nodes + not aborted + not non-retryable.
+        // Partial result (node_count > 0) is never retried.
+        let is_retryable = |o: &SubtaskOutcome| {
+            o.error.is_some() && o.node_count == 0 && !abort.is_set() && !non_retryable
+        };
+
+        // --- Attempt 2: reduced_complexity=true, minimal_skills=true ---
+        let final_outcome = if is_retryable(&outcome1) {
+            tracing::warn!(
+                subtask = %subtask.id,
+                error = outcome1.error.as_deref().unwrap_or(""),
+                "concurrent worker: subtask empty, retrying with minimal skills (attempt 2)"
+            );
+            run_subtask(subtask, plan, request, llm, sink, abort, true, true).await
+        } else {
+            outcome1
+        };
+
+        // Emit progress.
+        let zero = final_outcome.node_count == 0;
+        let node_count = final_outcome.node_count;
+        let err_msg = final_outcome.error.clone();
+
+        if zero {
+            progress_fn(Progress::SubtaskFailed {
+                id: subtask.id.clone(),
+                error: err_msg.unwrap_or_default(),
+            });
+        } else {
+            progress_fn(Progress::SubtaskDone {
+                id: subtask.id.clone(),
+                node_count,
+            });
+        }
+
+        outcomes.push(final_outcome);
+    }
+
+    outcomes
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -227,4 +410,327 @@ mod tests {
         assert_eq!(groups[1].screen, "settings");
         assert_eq!(groups[1].indices, vec![2]);
     }
+
+    // ── Task B1: BufferDocSink ────────────────────────────────────────────────
+
+    use crate::test_support::{ScriptResponse, ScriptedLlm};
+    use crate::types::LlmError;
+    use futures::executor::block_on;
+    use op_editor_core::EditorState;
+
+    fn make_plan_with_subtasks(subtask_ids: &[&str]) -> crate::plan::OrchestratorPlan {
+        use crate::plan::{OrchestratorPlan, Region, RootFrameSpec};
+        OrchestratorPlan {
+            root_frame: RootFrameSpec {
+                id: "root".into(),
+                name: "Page".into(),
+                width: 1200.0,
+                height: 800.0,
+                layout: None,
+                gap: None,
+                padding: None,
+                fill: None,
+            },
+            subtasks: subtask_ids
+                .iter()
+                .map(|id| crate::plan::Subtask {
+                    id: id.to_string(),
+                    label: id.to_string(),
+                    region: Region {
+                        width: 1200.0,
+                        height: 400.0,
+                    },
+                    id_prefix: id.to_string(),
+                    parent_frame_id: Some("root".into()),
+                    elements: None,
+                    screen: None,
+                })
+                .collect(),
+            style_guide_name: None,
+        }
+    }
+
+    fn make_req() -> crate::types::DesignRequest {
+        crate::types::DesignRequest {
+            prompt: "test".into(),
+            model: None,
+            provider: None,
+            design_md: None,
+            concurrency: 2,
+        }
+    }
+
+    const NODE_JSON: &str = r#"[{"type":"frame","id":"h1","name":"H","x":0,"y":0,"width":100,"height":100,"children":[]}]"#;
+
+    /// `BufferDocSink` collects commands without modifying a real doc.
+    #[test]
+    fn buffer_doc_sink_collects_commands() {
+        let mut sink = BufferDocSink::new(EditorState::new());
+        // apply() should always return true and buffer the command.
+        let applied = sink.apply(EditorCommand::InsertSubtree {
+            nodes: vec![],
+            parent_id: op_editor_core::NodeId::NONE,
+        });
+        assert!(applied, "BufferDocSink.apply must always return true");
+        assert_eq!(sink.commands.len(), 1);
+    }
+
+    /// `state()` on `BufferDocSink` returns the snapshot passed at construction.
+    #[test]
+    fn buffer_doc_sink_state_returns_snapshot() {
+        let state = EditorState::new();
+        let sink = BufferDocSink::new(state.clone());
+        // The returned reference should be valid (same content).
+        let _ = sink.state();
+    }
+
+    /// `BufferDocSink` tracks undo-batch depth correctly.
+    #[test]
+    fn buffer_doc_sink_undo_batch_depth() {
+        let mut sink = BufferDocSink::new(EditorState::new());
+        assert_eq!(sink.batch_depth, 0);
+        sink.begin_undo_batch();
+        assert_eq!(sink.batch_depth, 1);
+        sink.end_undo_batch();
+        assert_eq!(sink.batch_depth, 0);
+    }
+
+    // ── Task B1: run_screen_group_worker ─────────────────────────────────────
+
+    /// Happy path: both subtasks succeed on attempt 1.
+    #[test]
+    fn worker_happy_path_both_succeed() {
+        let plan = make_plan_with_subtasks(&["s0", "s1"]);
+        let req = make_req();
+        let llm = ScriptedLlm::new(vec![
+            ScriptResponse::Text(NODE_JSON.into()), // s0 attempt 1
+            ScriptResponse::Text(NODE_JSON.into()), // s1 attempt 1
+        ]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0, 1],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let mut events: Vec<Progress> = Vec::new();
+        let outcomes = block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |p| events.push(p),
+        ));
+        assert_eq!(outcomes.len(), 2);
+        assert_eq!(outcomes[0].node_count, 1);
+        assert_eq!(outcomes[1].node_count, 1);
+        // Buffer collected 2 InsertSubtree commands (one per subtask).
+        assert_eq!(sink.commands.len(), 2);
+    }
+
+    /// Attempt-1 zero-node + retryable → attempt 2 runs with (true, true).
+    /// We verify this by scripting: attempt 1 → error, attempt 2 → nodes.
+    #[test]
+    fn worker_attempt1_zero_retryable_triggers_attempt2() {
+        let plan = make_plan_with_subtasks(&["s0"]);
+        let req = make_req();
+        let llm = ScriptedLlm::new(vec![
+            // Attempt 1 — LLM error → zero nodes, retryable
+            ScriptResponse::Fail(LlmError {
+                message: "timeout".into(),
+                aborted: false,
+            }),
+            // Attempt 2 — succeeds
+            ScriptResponse::Text(NODE_JSON.into()),
+        ]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let mut events: Vec<Progress> = Vec::new();
+        let outcomes = block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |p| events.push(p),
+        ));
+        assert_eq!(outcomes.len(), 1);
+        // Final outcome should show nodes from attempt 2.
+        assert_eq!(
+            outcomes[0].node_count, 1,
+            "attempt 2 should have produced a node"
+        );
+        assert_eq!(sink.commands.len(), 1, "one InsertSubtree from attempt 2");
+    }
+
+    /// Non-retryable error on attempt 1 → NO attempt 2.
+    #[test]
+    fn worker_non_retryable_error_no_retry() {
+        let plan = make_plan_with_subtasks(&["s0"]);
+        let req = make_req();
+        let llm = ScriptedLlm::new(vec![
+            // Attempt 1 — non-retryable HTTP 401
+            ScriptResponse::Fail(LlmError {
+                message: "HTTP 401 Unauthorized".into(),
+                aborted: false,
+            }),
+            // Attempt 2 — should NOT be called; if it is, the test fails by
+            // observing the scripted LLM's exhausted error.
+        ]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let mut events: Vec<Progress> = Vec::new();
+        let outcomes = block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |p| events.push(p),
+        ));
+        assert_eq!(outcomes.len(), 1);
+        assert_eq!(outcomes[0].node_count, 0, "non-retryable must stay zero");
+        // Verify the error message is the original 401, not "scripted LLM exhausted"
+        assert!(
+            outcomes[0].error.as_deref().unwrap_or("").contains("401"),
+            "error should carry original 401 message"
+        );
+        // No InsertSubtree in the buffer (nothing was applied).
+        assert_eq!(sink.commands.len(), 0);
+    }
+
+    /// Partial result (node_count > 0) on attempt 1 → never retried.
+    #[test]
+    fn worker_partial_result_never_retried() {
+        let plan = make_plan_with_subtasks(&["s0"]);
+        let req = make_req();
+        // Attempt 1 produces nodes — the scripted LLM has only one response.
+        let llm = ScriptedLlm::new(vec![ScriptResponse::Text(NODE_JSON.into())]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let outcomes = block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |_| {},
+        ));
+        assert_eq!(outcomes.len(), 1);
+        assert_eq!(outcomes[0].node_count, 1);
+        // Only 1 command buffered (attempt 1 only).
+        assert_eq!(sink.commands.len(), 1);
+    }
+
+    /// Abort flag set before the loop → worker returns immediately with no outcomes.
+    #[test]
+    fn worker_aborts_at_loop_head() {
+        let plan = make_plan_with_subtasks(&["s0", "s1"]);
+        let req = make_req();
+        let llm = ScriptedLlm::new(vec![]); // should not be called
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0, 1],
+        };
+        let abort = AbortFlag::new();
+        abort.set(); // set before the worker runs
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let outcomes = block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |_| {},
+        ));
+        assert!(
+            outcomes.is_empty(),
+            "aborted before loop: no outcomes expected"
+        );
+        assert_eq!(sink.commands.len(), 0);
+    }
+
+    /// Progress events: SubtaskStarted is emitted before each subtask;
+    /// SubtaskDone emitted for successful subtasks.
+    #[test]
+    fn worker_emits_started_and_done_progress() {
+        let plan = make_plan_with_subtasks(&["s0", "s1"]);
+        let req = make_req();
+        let llm = ScriptedLlm::new(vec![
+            ScriptResponse::Text(NODE_JSON.into()),
+            ScriptResponse::Text(NODE_JSON.into()),
+        ]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0, 1],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let mut events: Vec<String> = Vec::new();
+        block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |p| match &p {
+                Progress::SubtaskStarted { id, .. } => events.push(format!("start:{id}")),
+                Progress::SubtaskDone { id, .. } => events.push(format!("done:{id}")),
+                Progress::SubtaskFailed { id, .. } => events.push(format!("fail:{id}")),
+                _ => {}
+            },
+        ));
+        assert_eq!(events, vec!["start:s0", "done:s0", "start:s1", "done:s1"],);
+    }
+
+    /// SubtaskFailed progress emitted when final outcome is zero nodes.
+    #[test]
+    fn worker_emits_failed_progress_on_zero_nodes() {
+        let plan = make_plan_with_subtasks(&["s0"]);
+        let req = make_req();
+        // Both attempts produce garbage → zero nodes.
+        let llm = ScriptedLlm::new(vec![
+            ScriptResponse::Text("not json".into()),
+            ScriptResponse::Text("also not json".into()),
+        ]);
+        let group = ScreenGroup {
+            screen: "home".into(),
+            indices: vec![0],
+        };
+        let abort = AbortFlag::new();
+        let mut sink = BufferDocSink::new(EditorState::new());
+        let mut events: Vec<String> = Vec::new();
+        block_on(run_screen_group_worker(
+            &group,
+            &plan,
+            &req,
+            &llm,
+            &abort,
+            &mut sink,
+            &mut |p| {
+                if let Progress::SubtaskFailed { id, .. } = &p {
+                    events.push(format!("fail:{id}"));
+                }
+            },
+        ));
+        assert_eq!(events, vec!["fail:s0"]);
+    }
 }