Bug: orchestration follow-up prompt is not reliably inserted for inherited child agents #253

Problem

When using orchestration agents with inheritParentContext: true, a follow-up prompt can fail to appear as a new instruction in the child agent. In the observed workflow, we attempted to redirect/continue an orchestrated child, but the new prompt did not get inserted into the child session.

This is especially dangerous for inherited children: the child starts from a duplicated parent transcript, so without an explicit inserted handoff/follow-up prompt it may continue interpreting the inherited conversation as if it is still the parent/orchestrator. That breaks role separation and makes parallel orchestration unreliable.

Why this matters

The existing bootstrap wrapper in src/mcp/shared/orchestrationPrompt.ts is the right idea: provider-visible text is the only reliable cross-provider instruction channel. The child must see, in its own transcript/composer input, that inherited history is read-only background and that the newly supplied task is its active instruction.

If orchestration_send_prompt reports success but the prompt is not actually inserted/submitted, the parent believes the child was redirected while the child keeps operating from stale inherited context.

Suspected areas

• src/mcp/runtime/createBuiltInMcpServer.ts around orchestration_send_prompt
• submitPrompt(...) behavior for Claude/Codex sessions
• OrchestrationBridge.notePromptSubmitted(...) and lifecycle reporting after prompt delivery
• markBootstrapPromptDelivered(...) ordering and whether bootstrap delivery is marked before/after real insertion succeeds
• renderer/session readiness path used by orchestration prompt submission

Expected behavior

• orchestration_send_prompt should only return ok: true after the prompt has been inserted/submitted to the target child session.
• For inherited children whose bootstrap handoff has not been delivered, the first prompt must be wrapped with buildOrchestrationBootstrapPrompt(...).
• If insertion/submission fails, the MCP tool must return a clear failure with a stable error code and message.
• Prompt delivery metadata should not advance in a way that makes wait_agents or status surfaces believe a prompt is in flight unless the prompt was actually accepted by the target session.
• The child transcript/feed should contain the inserted prompt so the child understands it is an orchestrated child, not the original parent from the inherited transcript.

Acceptance criteria

• Add a regression test or harness coverage for orchestration_send_prompt against an inherited child.
• Prove the first follow-up prompt to an inherited child is wrapped and inserted exactly once.
• Prove a failed prompt insertion returns ok: false and does not mark bootstrap delivery or prompt submission state.
• Prove a successful prompt insertion updates lifecycle/status so the parent can wait on the child without seeing stale completion.
• Include a WHY comment near the delivery/order check explaining that inherited children can otherwise misidentify themselves as the parent agent.

Related context

This is separate from Root Access MCP (#242). Root Access is about global/non-descendant control. This bug is about scoped orchestration prompt delivery to a valid child session and the correctness of inherited-context role handoff.