Summary

This PR restores the work from #18493 after the original fork branch was deleted and GitHub could not reopen the closed pull request.

This PR adds agent_control: a durable orchestration layer that lets Hermes profiles manage other Hermes profiles as persistent peer agents over ACP.

The core idea is simple: delegate_task is for disposable subagents, Kanban is for asynchronous task-board workflows, and agent_control is for direct manager/team orchestration. A Hermes orchestrator can now command named specialist profiles with their own memory, tools, skills, config, and resumable session history.

This turns Hermes profiles into composable agent teammates without inventing a separate swarm runtime. It reuses the primitives Hermes already trusts: profiles for identity, ACP for process/session control, toolsets for capability boundaries, and SQLite for durable state.

This builds on the external orchestration direction from #14009, but moves the control plane into Hermes as an opt-in toolset designed for production reliability.

What changed

• Add a lightweight ACP stdio JSON-RPC client for Hermes profile sessions.
• Add a durable SQLite-backed control plane for profile-agent handles, runs, and leases.
• Add the agent_control toolset and tools:
  • agent_start
  • agent_prompt
  • agent_status
  • agent_list
  • agent_fork
• Keep editor ACP behavior unchanged by default.
• Allow agent-control ACP subprocesses to use the controlled profile's CLI toolsets.
• Prevent delegate_task children from inheriting agent_control.
• Keep dangerous permission approval policy outside the model-facing schema; trusted local automation can opt into one-shot approvals only via process/admin configuration.
• Omit model-facing cancellation until Hermes has an async or pollable run API with ACP response dispatching strong enough to cancel in-flight work reliably.
• Add user docs and an ADR comparing delegate_task, agent_control, and Kanban.
• Convert the existing Kanban architecture diagram to Mermaid so docs diagram lint stays green when website docs are touched.

Reliability notes

• SQLite session leases are keyed by (profile, session_id), so duplicate handles or separate orchestrator profiles cannot race the same ACP session history.
• agent_fork takes the same source-session lease before branching history.
• Expired leases are marked as errors instead of leaving handles stuck.
• Existing ACP sessions are loaded strictly; missing sessions fail instead of silently creating ghost state.
• Dangerous permission requests fail closed by default.
• The model cannot set approval_policy="allow_once" through tool arguments; permissive behavior requires HERMES_AGENT_CONTROL_APPROVAL_POLICY=allow_once.
• agent_control is gated on the ACP extra.

Tests

• .venv/bin/python -m pytest tests/acp/test_client.py tests/agent/test_orchestration.py tests/acp/test_session.py tests/tools/test_delegate.py tests/tools/test_registry.py::TestBuiltinDiscovery::test_matches_previous_manual_builtin_tool_set -q
  • 190 passed
• .venv/bin/python -m compileall acp_adapter/client.py acp_adapter/session.py agent/orchestration.py tools/agent_control_tool.py hermes_cli/tools_config.py tests/acp/test_client.py tests/acp/test_session.py tests/agent/test_orchestration.py tests/tools/test_delegate.py tests/tools/test_registry.py
• agent_control registry smoke check
• agent_control toolset smoke check
• .venv/bin/ascii-guard lint website/docs
• npm run build
  • completed successfully with the venv Python on PATH so the skill metadata prebuild matches CI's pyyaml setup.
• git diff --check

The targeted test set covers the new ACP client, the durable orchestration control plane, ACP toolset behavior, and the regression that prevents delegate_task children from inheriting agent_control. The full repository suite is left to CI.