Summary

Hi OpenClaw team, thanks for the great project.

After upgrading to v2026.2.19-2, we observed that subagent completion often goes directly to requester chat, which breaks manager-first multi-agent workflows. If you're in a hurry, please jump directly to the mermaid chart below.

Problem

In manager -> child -> grandchild orchestration, completion delivery now tends to bypass parent aggregation and appears directly in user chat.

This causes two major problems:

1. Hierarchy value is degraded

   • Manager cannot reliably aggregate child output first.
   • Child cannot reliably aggregate grandchild output first.
   • Parent-level validation/synthesis step is skipped.

2. User chat is flooded

   • Raw worker outputs flood end-user chat.
   • UX becomes noisy and hard to use in real multi-agent tasks.

Additional impact:

• Harder to enforce dedupe/quality/redaction before user delivery.
• Parallel completions can appear out of order.
• Self-eval style pipelines lose “manager as convergence point”.

Real workflow impact (Self-Eval example)

Our self-eval workflow depends on hierarchy (operator/reviewer/executor style).
Current behavior pushes worker-level completions directly to user chat, which reduces orchestration quality and final report control.

Request

Please provide a configurable completion delivery mode, e.g.:

• manager-first (child -> parent -> user)
• direct (current behavior)

Or expose the completion routing behavior as a config/tool option so orchestration users can restore manager-first flow.

There are prior announce-related issues, but this report is specifically about v2026.2.19-2 completion routing semantics.

Steps to reproduce

Minimal repro

1. Spawn sub0 via sessions_spawn.
2. Ask sub0 to spawn sub1 and have sub1 reply OK_SUB1_ROUTE_TEST.
3. Observe both SUB0_SPAWNED and OK_SUB1_ROUTE_TEST appear directly in requester chat.

Expected behavior

Subagent results should flow upward through the agent hierarchy first (grandchild → child → manager).

The manager should consolidate/validate child outputs and send one curated user-facing response.

User chat should not receive raw completion messages from child/grandchild agents by default.

Expected vs current

flowchart TD
    U[User Chat] --> M[Manager Agent]

    subgraph expected["Expected: manager-first"]
      M --> C[Child Subagent]
      C --> G[Grandchild Subagent]
      G --> C
      C --> M
      M --> U
    end

    subgraph current["Current in v2026.2.19-2: direct completion"]
      M -.spawn.-> C2[Child Subagent]
      C2 -.spawn.-> G2[Grandchild Subagent]
      C2 --> U
      G2 --> U
    end

Actual behavior

In v2026.2.19-2, completion messages from child/grandchild agents are delivered directly to the requester chat.

Parent agents do not reliably receive descendant completion as internal input first.

This bypasses manager-side aggregation and causes user chat noise/flooding in multi-agent workflows.

OpenClaw version

v2026.2.19-2

Operating system

Mac 26.2/Ubuntu 22

Install method

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response