Summary
This issue tracks the phased rollout of a model-agnostic, policy-governed agent harness across S-CORE repositories. It covers shared control-plane concerns only: run contract, CI orchestration, policy-as-code, evidence schema, onboarding waves, and hard-fail governance.
Domain-specific harness work (task corpora, consistency rules, role models, outer-loop optimization) lives in two separate domain parent issues linked below.
Why
Agent-assisted changes need to be reproducible, policy-governed, and auditable across multiple repositories. The docs-as-code traceability metrics and gate are the reference pattern. This rollout scales the same control-plane model to the wider portfolio while keeping domain-specific complexity out of the shared infrastructure.
Lane Model
- Lane A (mandatory): all merge-critical checks run with open-source tooling only. Merge eligibility is determined exclusively from Lane A results.
- Lane B (optional accelerator): proprietary or LLM-assisted helpers may improve productivity but cannot be required for merge and cannot override a Lane A denial.
Scope (shared control-plane only)
- One common agent run contract across repos
- Shared CI workflow templates and required check names
- Policy-as-code bundle for merge decisions
- Standardized evidence artifact schema (base fields + domain extensions)
- Wave-based onboarding across repo groups
- Hard-fail governance after domain-specific readiness criteria are met
Out of Scope for this issue
- Domain-specific task corpora, consistency rules, and harness interfaces (see domain parent issues)
- Internal OEM benchmark analytics, release authority, and waiver logic
- Rewriting existing module architectures
Deliverables
- Reusable run contract template
- Shared CI workflow with required artifact outputs
- Policy bundle with allow/deny decision format
- Evidence artifact schema v1 with domain extension mechanism
- Wave onboarding tracker
- Hard-fail precondition checklist per domain
- Short repository entrypoint guidance (
AGENTS.md/equivalent map) that links to indexed domain docs instead of duplicating them
- Pilot-ready conventions for cheap candidate validation and index-first run-history navigation
Acceptance Criteria
- All Wave 1 repositories produce standardized Lane A evidence artifacts.
- Policy gate blocks seeded regressions in pilot repos.
- Lane A reruns are deterministic within agreed tolerance.
- Hard-fail is enabled only after domain-specific readiness criteria are met.
- Shared repo guidance remains concise and navigational; domain detail lives in indexed, queryable subsystem docs.
- Pilot domains can validate malformed candidates cheaply before expensive runs and expose a small summary-first trace navigation surface.
Sub-issues
Shared control-plane
Domain parents
Suggested Labels
community:infrastructure, automation, safety, testing, software_development_process
Summary
This issue tracks the phased rollout of a model-agnostic, policy-governed agent harness across S-CORE repositories. It covers shared control-plane concerns only: run contract, CI orchestration, policy-as-code, evidence schema, onboarding waves, and hard-fail governance.
Domain-specific harness work (task corpora, consistency rules, role models, outer-loop optimization) lives in two separate domain parent issues linked below.
Why
Agent-assisted changes need to be reproducible, policy-governed, and auditable across multiple repositories. The docs-as-code traceability metrics and gate are the reference pattern. This rollout scales the same control-plane model to the wider portfolio while keeping domain-specific complexity out of the shared infrastructure.
Lane Model
Scope (shared control-plane only)
Out of Scope for this issue
Deliverables
AGENTS.md/equivalent map) that links to indexed domain docs instead of duplicating themAcceptance Criteria
Sub-issues
Shared control-plane
Domain parents
Suggested Labels
community:infrastructure, automation, safety, testing, software_development_process