Environment

• OpenClaw Gateway: 2026.6.5
• oh-my-openclaw: v0.21.3
• Model: deepseek-art/deepseek-v4-pro
• Scenario: Delegate persona using sessions_yield to orchestrate subagents

Problem

When a parent agent uses sessions_yield to wait for subagent completions, the completion is delivered by injecting it as passive prompt context into the parent agent's next turn. The parent agent is expected to read this context, recognize that a subagent has completed, and take appropriate action (verify results → update plan → continue or finish).

However, LLMs inherently treat prompt context as reference material that may be acted upon or ignored — there is no mechanism forcing the parent agent to process the completion event. Whether the parent agent picks up the completion and continues the orchestration loop, or ignores it and produces unrelated output, is entirely determined by the LLM's non-deterministic behavior on that specific turn.

This means a multi-agent orchestration loop (subagent-complete → parent-review → continue/end) can silently break at any turn, with no error, no recovery path, and no indication that anything went wrong.

Diagnostic Evidence

We injected 5 console.log diagnostic points into sendSubagentAnnounceDirectly (in subagent-announce-delivery module) to trace the delivery path:

Log from a subagent completion event:

12:51:04 subagent_announce_active_check
 isActive: false ← correct after sessions_yield
 sessionId: 039913d5-...

12:51:04 subagent_announce_direct_delivery
 shouldDeliver: false ← dashboard session, no external channel
 hasChannel: false

12:51:10 subagent_announce_result
 path: "direct" ← correct path taken
 delivered: true ← delivery succeeded

The delivery path is correct: sessions_yield correctly sets the parent session to non-active, the direct delivery path is taken, and delivered: true confirms the completion message was injected into the parent session's prompt context.

Observed parent agent behavior (3 attempts, same scenario):

Same completion format, same Delegate persona, same Gateway configuration — three different outcomes, two of which silently broke the orchestration workflow.

Root Cause

The root cause is the delivery mechanism itself, not prompt wording or instruction conflicts. Subagent completion is delivered as passive prompt context, and LLMs inherently treat all context as optional — reference material they may read and act on, or skip and ignore.

For a critical workflow step like "subagent reports completion to parent", a passive context injection is too weak. The parent agent needs a stronger harness that forces it to process the completion result rather than offering it as one of many context blocks the LLM can overlook.

The problem is not contradictory instructions — it's that no instruction in prompt context can guarantee execution by an LLM. Context is, by definition, advisory.