Problem

Sub-agents spawned via spawn_agent do not receive the embedded AGENTS.md operating rules layer that main-agent sessions get. They only receive their own markdown body, optional project instructions (from the workspace directory), and the hardcoded HeadlessExecutionContract.

This means sub-agents operate without the core safety guardrails that main agents inherit:

• Grounding rules (e.g., "Never state runtime facts without checking with a tool")
• Search/citation policy (inline hyperlink requirements, when to search vs. answer)
• Safety boundaries (Resource Hard Deny awareness, "I don't know beats a confident wrong answer")
• Proactive check-back expectations
• Search decision rules

Sub-agents are autonomous, non-interactive workers. Running them without these guardrails is worse than running main agents without them — the sub-agent is acting on behalf of the parent session with no approval loop to catch misbehavior.

Current sub-agent prompt stack

Currently assembled in SubAgentActor.BuildSystemPrompt():

[Sub-Agent Markdown Body]
[Skill Overlay (if routed via metadata.subagent)]
[HeadlessExecutionContract]

Missing entirely:

• Embedded AGENTS.md (operating rules, safety, grounding)
• SOUL.md (personality/tone)
• TOOLING.md (host capabilities)

Project instructions (AGENTS.md/CLAUDE.md/CONTEXT.md from workspace) do load correctly via ResolveProjectInstructions() in SubAgentSpawner.cs.

Proposed prompt stack

[Embedded AGENTS.md]           — base operating rules, safety, grounding
[SOUL.md]                      — tone/communication style (optional)
[TOOLING.md]                   — host capabilities (likely redundant, tool index already available)
[Project instructions]         — workspace/domain context (already works)
[Sub-agent Markdown Body]      — role definition, what it does
[Skill Overlay]                — skill-specific context if routed
[HeadlessExecutionContract]    — execution contract, always at the bottom

Implementation sketch

In SubAgentActor.BuildSystemPrompt(), instead of:

var basePrompt = string.IsNullOrWhiteSpace(definition.ProjectInstructions)
    ? definition.SystemPrompt
    : SystemPromptAssembler.Assemble(
        agents: definition.SystemPrompt,
        projectInstructions: definition.ProjectInstructions);

Pass the embedded AGENTS.md and other identity layers through SystemPromptAssembler.Assemble(), then append the sub-agent markdown body, then the HeadlessExecutionContract.

The embedded AGENTS.md can be read from CachedAgents.Value (same source used by FileSystemPromptProvider), and project instructions already resolve from definition.ProjectInstructions.

Code locations

• SubAgentActor.BuildSystemPrompt() — prompt assembly (main change)
• SubAgentSpawner.ResolveProjectInstructions() — already works, no change needed
• SubAgentDefinition — may need additional fields or a different assembly approach
• FileSystemPromptProvider.GetSystemPrompt() — reference for how main agents assemble their prompt

Notes

• The HeadlessExecutionContract is currently hardcoded in SubAgentActor and should remain appended at the end
• SOUL.md and TOOLING.md may be worth including but are lower priority — the critical missing layer is embedded AGENTS.md
• This is a prompt-layering change only, no API or protocol changes