Sub-agent orchestration

[bug-ops]

Overview

Implement sub-agent orchestration for Zeph, enabling the main agent to delegate tasks to specialized child agents running in isolated contexts with controlled permissions. Sub-agents communicate with the main agent via A2A protocol semantics.

Security model: Zero-Trust — every sub-agent starts with zero permissions. All grants (secrets, tools) require explicit user approval, are TTL-bounded, and auto-revoked on expiry or sub-agent termination. Secrets never persist to disk or message history.

Architecture document: .local/plan/sub-agent-orchestration.md

Design Principles

1. Zero-trust: every sub-agent is untrusted by default
2. Least privilege: grant only minimum required for the specific task
3. Time-bounded access: all grants have TTL, auto-revoked on expiry
4. No nesting: sub-agents cannot spawn sub-agents
5. Main agent is the single user-facing interface
6. A2A protocol for all orchestrator↔sub-agent communication

Phases

• M28-A: Sub-agent definition format and parser #710 M28-A: Sub-agent definition format and parser
• M28-B: SubAgentManager core with in-process A2A channels #711 M28-B: SubAgentManager core with in-process A2A channels
• M28-C: Tool and skill filtering for sub-agents #712 M28-C: Tool and skill filtering for sub-agents
• M28-D: Permission and secret delegation protocol #713 M28-D: Permission and secret delegation protocol (zero-trust)
• M28-E: Background sub-agent execution #714 M28-E: Background sub-agent execution
• M28-F: TUI sub-agent panel #715 M28-F: TUI sub-agent panel
• M28-G: CLI /agent commands #716 M28-G: CLI /agent commands