IDKMesh

Enterprise Control Plane Architecture

Status: proposed enterprise baseline
Date: 2026-09-22
Program: issue #667
Scope: enterprise deployment, trust boundaries, tenancy, identity, data governance, auditability, resilience, and supply-chain controls.

1. Decision in one sentence

IDKMesh keeps the GitHub-first G0/G1 profile as the default product experience, and adds enterprise controls as a strict overlay that can graduate to a dedicated G2/G3 control service only when tenancy, private-network, audit, state, latency, or organizational policy requires it.

Enterprise readiness is therefore not synonymous with Kubernetes, a central database, or a new login system. It means that every privileged action has a bounded authority, every tenant/data boundary is explicit, every important decision is attributable and recoverable, and operational failure has a tested safe state.

2. What enterprise means here

An enterprise deployment must be able to answer these questions from evidence:

  1. Which tenant/project does this action belong to?
  2. Which human or service identity initiated it?
  3. Which policy version authorized, denied, or escalated it?
  4. Which exact repository/resource revision did it operate on?
  5. Which data classification and egress rules applied?
  6. Which secret/workload identity was available, and only after which gate?
  7. What worker/provider actually executed the work?
  8. Which independent evidence was produced?
  9. Who had integration authority?
  10. Can the control state be reconstructed after a runner/service failure?
  11. Can an auditor detect rewriting or deletion of privileged events?
  12. Can the operator restore service/state within the declared recovery objectives?

A model response, a worker claim, a successful CI job, an administrator identity, or a vendor badge is not by itself an answer to these questions.

3. Deployment progression

The existing G0-G4 progression remains authoritative.

G0 — repository control plane

GitHub repository + Actions + hosted connectors.

Best for a single project and normal asynchronous software work. No project-owned always-on service is required.

G1 — repository control plane plus optional nodes

G0 plus bounded local/private workers such as a workstation, GPU node, or private-network execution node.

The node remains a worker. It is not the canonical scheduler, database, verifier, or integration authority.

G2 — dedicated control service

Use when one organization/project needs requirements that G0/G1 cannot satisfy cleanly, such as:

G2 should normally be single-organization or dedicated tenancy.

G3 — shared multi-project / multi-tenant control plane

Use only when one service deliberately manages multiple projects, repositories, or organizational tenants.

G3 adds a hard tenant boundary to every API request, state key, queue item, cache entry, audit event, secret binding, connector binding, and idempotency key.

G4 — federation

Multiple independently operated control planes exchange bounded work/evidence. This remains future research and must not weaken the enterprise isolation requirements of G2/G3.

4. Enterprise reference topology

Human users / enterprise IdP
        |
        | SSO / GitHub identity / OIDC
        v
+------------------------------+
| GitHub organization/repos    |
| code, issues, PRs, rulesets  |
| protected integration        |
+---------------+--------------+
                |
                | webhooks / Actions / GitHub App
                v
+----------------------------------------------------+
| Optional IDKMesh G2/G3 control service            |
|                                                    |
| API / ingress plane                               |
|   -> authentication                               |
|   -> tenant context                               |
|   -> authorization + policy decision              |
|                                                    |
| coordination plane                                |
|   -> WorkUnit admission                           |
|   -> claims / leases / idempotency                |
|   -> connector routing                            |
|   -> run state                                    |
|                                                    |
| evidence / audit plane                            |
|   -> append-only audit events                     |
|   -> evidence references / digests                |
|   -> SIEM/archive export                          |
|                                                    |
| state plane                                       |
|   -> durable DB / ledger                          |
|   -> encrypted backups                            |
+---------+------------------+-----------------------+
          |                  |
          | bounded dispatch | bounded dispatch
          v                  v
 hosted providers       private/local workers
          |                  |
          +--------+---------+
                   |
                   v
          candidate + evidence
                   |
                   v
             GitHub PR/checks
                   |
                   v
          protected human integration

GitHub remains the canonical code and integration authority unless a future ADR explicitly changes that boundary.

5. Four planes, four authorities

Enterprise deployments must not collapse these planes into one privileged service account.

5.1 Source/integration plane

Owns repositories, protected branches, releases, rulesets, and final integration.

A worker/provider does not receive this authority merely because it can create a candidate branch or pull request.

5.2 Control plane

Owns admission, routing, claims, run state, retry/idempotency, policy evaluation, and operator status.

The control plane may decide that work is eligible to execute. It does not decide that the resulting candidate is correct or merged.

5.3 Execution plane

Runs untrusted or semi-trusted work in hosted providers, CI sandboxes, local nodes, private runners, containers, VMs, or other substrates.

Execution identity is provenance, not repository authority.

5.4 Evidence/audit plane

Records verification evidence and privileged control events.

Evidence can recommend or document a decision. The audit stream records what happened. Neither one is permission to integrate.

6. Tenant boundary

For G3 and any deployment claiming shared tenancy:

Issue #669 owns executable isolation work.

7. Identity and authority

The normalized identity model should distinguish at least:

Enterprise authorization is stage-specific.

actor + tenant + action + resource + risk + data class + policy revision
                            |
                            v
                 allow / deny / approval required

High-risk operations should support separation of duties. A single person may hold multiple organizational roles, but enterprise policy can require distinct actors for a particular execution/integration path.

Break-glass access must be explicit, time bounded, reason required, and audited. It must not silently become the normal operating path.

Issue #670 owns the identity/RBAC/ABAC/SoD implementation.

8. Data classification and egress

The enterprise baseline uses four minimum data classes:

A WorkUnit or project policy can tighten classification. An agent/model cannot lower it.

External processing is a non-compensating eligibility check. If policy forbids a data class from leaving the approved trust boundary, model quality, cost, speed, or administrator convenience cannot override that rule.

Enterprise workers should support:

Issue #672 owns the data/egress/secrets boundary.

9. Secret and workload identity

Tracked configuration stores references, never credential values.

Preferred order for enterprise/cloud access:

  1. short-lived workload identity / OIDC;
  2. enterprise secret manager reference;
  3. GitHub Environment/repository secret where appropriate;
  4. long-lived cloud access keys only where policy explicitly permits them.

Authorization happens before secret materialization.

Secret access should be auditable as a decision and resource reference without recording the secret value.

10. Audit model

Ordinary application logs are not the enterprise audit ledger.

A privileged audit event should bind:

The baseline requires append-only semantics. G2/G3 production deployments should support export to an organization-controlled archive/SIEM.

Audit evidence records decisions; it does not grant authority by existing.

Issue #671 owns the audit ledger/export contract.

11. Reliability and recovery

Enterprise operation requires explicit objectives rather than adjectives such as “high availability.”

The Enterprise Control Profile records:

These are declared objectives until measured by a pilot.

Required failure behavior:

Issue #673 owns operational SLO/DR work.

12. Supply chain

Enterprise release controls should progressively provide:

GitHub artifact attestations and OIDC can support these controls, but an attestation proves provenance/integrity facts, not functional correctness.

Issue #674 owns release/supply-chain hardening.

13. Observability

Operational telemetry and audit evidence are separate.

Metrics/logs/traces should support:

Telemetry must avoid secret values and should minimize sensitive task content.

14. Change and incident management

Enterprise operation needs an accountable control loop:

proposed change
 -> reviewed policy/config/code
 -> exact revision
 -> staged validation
 -> deployment
 -> observed health
 -> rollback or acceptance

Security incidents should have stable incident identifiers that can be referenced by audit events, emergency changes, and post-incident evidence.

Emergency access must leave more evidence than normal access, not less.

15. G2/G3 service requirements

The optional service tracked by #675 should be introduced only after the repository-only pilot identifies a need or an enterprise control explicitly requires it.

Minimum service qualities:

A database, queue, container orchestrator, or cloud vendor is an implementation choice beneath these invariants.

16. Conformance, not certification

IDKMesh may map controls to common enterprise frameworks in future, but the project MUST distinguish:

A passing IDKMesh enterprise preflight is not a SOC 2, ISO 27001, FedRAMP, HIPAA, PCI DSS, or similar certification claim.

Issue #676 owns destructive conformance testing.

17. Current implementation boundary

Implemented on main today:

Planned/open enterprise work:

Do not describe the planned controls as production-enforced until their issues and acceptance evidence are complete.

18. Enterprise success gate

The baseline is ready for an enterprise pilot only when the system can demonstrate that:

That evidence, rather than infrastructure size, is what makes the system enterprise-operable.