Skip to content

Persistent canonical role sessions with governed skill policy #15153

@stefanpieter

Description

@stefanpieter

Problem

Hermes can currently execute delegated work, but role/team workflows need first-class persistent role instances rather than throwaway subagents or artifact-only metadata.

For long-running product/engineering workflows, the desired behavior is:

  • stable canonical role identities, e.g. Lead / PM, Planner, Developer, Validator, Release Manager;
  • real persisted role sessions with durable history;
  • role isolation by plan/work item and canonical role;
  • the ability to retire/reopen a role session without losing history;
  • Lead-mediated findings ingestion instead of automatic remediation loops;
  • visible role runtime metadata for dashboards/API consumers;
  • explicit per-role skill policy so role behavior is consistent and evolvable.

Proposed direction

Add a persistent role-team runtime that treats canonical roles as durable SessionDB-backed role instances. Roles should be resumable by stable (plan_id, canonical_role) identity and expose machine-readable runtime/skill evidence in execution artifacts.

Skill policy should be a governed asset rather than prompt drift:

  • role-required skills;
  • role-recommended skills;
  • domain-triggered specialist skills;
  • compliance metadata showing whether required skills were loaded.

Prototype branch

A prototype branch is available here:

  • stefanpieter/hermes-agent:role-team-runtime-overhaul

Key commits:

  • de01b531 Implement persistent role team runtime
  • de836dd4 Add role skill policy loading
  • ba871923 chore: preserve local Hermes runtime customizations

Validation performed locally

Focused runtime / role skill validation:

venv/bin/python -m pytest tests/tools/test_role_invocation_tool.py tests/agent/test_role_skill_policy.py tests/agent/test_role_sessions.py -q
25 passed

Auxiliary/Codex/config local-preservation validation:

venv/bin/python -m pytest tests/agent/test_auxiliary_client.py tests/hermes_cli/test_codex_models.py tests/hermes_cli/test_managed_installs.py -q
115 passed

Additional checks:

venv/bin/python -m py_compile agent/auxiliary_client.py hermes_cli/codex_models.py hermes_cli/config.py hermes_cli/main.py tools/role_invocation_tool.py agent/role_runtime.py web/scripts/generate_org_chart_data.py
git diff --cached --check

Independent review was run on the role-runtime and skill-loading slices, and a final independent code review of the local-preservation code diff returned no blocking security or logic issues.

Notes

This issue is intended to preserve and discuss the architecture before/alongside PR review, especially because the role runtime is a workflow-level design change rather than a small isolated bug fix.

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3Low — cosmetic, nice to havecomp/agentCore agent loop, run_agent.py, prompt buildertype/featureNew feature or request

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions