Universal Core From The Uploaded Harness Zip

Bottom line

The universal value of the uploaded zip is the agentic SDLC operating model. Keep the structure, gates, run packets, front doors, and answer contracts. Move product names, deployment rules, team owners, validation commands, and red-zone specifics into a generated project adapter.

Keep in the universal core

1. Commissioning interview

   • Ask the human/operator for facts a code graph cannot infer.
   • Capture source-of-truth order, approval owners, answer style, production definition, Red Zone boundaries, and validation expectations.

2. Project adapter schema

   • Store generated repo/team facts in project-adapter.json and project.yaml.
   • Agents read this before touching code.

3. Agent front doors

   • Generate AGENTS.md for Codex and general coding agents.
   • Generate CLAUDE.md or slash-command templates for Claude Code.
   • Keep the front doors thin: point to the adapter, router, lanes, gates, and run packet.

4. Router and lanes

   • Keep 00_MAP.md and CONTEXT.md as the default map/router pattern.
   • Generate only the lanes a repo needs.
   • Default lane set: engineering-default, incidents, data, infra/provider-config, security, docs-product.

5. Stage flow

   • Current v0.8 execution sequence:

     intake -> workload classification + Layer 0 -> route -> code intelligence + design -> implement -> Red Zone -> QA -> prove -> live smoke -> self-heal -> signed review -> run packet -> handoff
     

   • Every stage either writes an artifact or validates one.

6. Run packet model

   • Keep a durable folder/file structure for every non-trivial run.
   • Required artifact families:
     • run/intake.json
     • run/workload-classification.json
     • run/route.json
     • foundation/assessment.json
     • goal/goal.json
     • context/manifest.json
     • ai/assurance.json
     • domain/assurance.json
     • evals/results.json plus its typed context-arm result files
     • trajectory/trajectory.json
     • waivers/waivers.json
     • design/anchors.json
     • session/events.jsonl
     • approvals/redzone.json
     • proof/proof.json
     • proof/portable.json
     • rca/rca.json when corrective-work RCA is required or supplied
     • review/review.json using the signed four-role valdris.review.v2 contract
     • run/packet.json using valdris.run-packet.v3 with an explicit assurance level and catalog snapshots; v2 is historical structural evidence only
     • handoff/final.md

7. Mechanical gates

   • Proof gate: no proof artifact, no done.
   • Layer 0 gate: no workload classification and required foundation assurance, no implementation.
   • Red Zone gate: no explicit approval, no risky mutation.
   • RCA gate: no runtime evidence, no confirmed cause claim.
   • Anchor gate: no fresh code anchors, no architecture claim.
   • Enterprise/AI aggregate gate: no coherent goal, context, control, eval, trajectory, and waiver state, no completion.
   • Review and run-packet gates: no authorized four-role signature and digest-bound final packet, no completion.
   • Smoke/deploy/migration gates as optional project modules.

8. Answer contract

   • Final answers stay decision-packet shaped:
     • Bottom line
     • Why
     • Proof
     • Fix/Plan
     • Your call

9. Connector event contract

   • Agents emit events like:
     • run.created
     • agent.connected
     • node.entered
     • gate.fired
     • artifact.written
     • approval.requested
     • run.blocked
     • run.completed
   • The visualizer consumes events. The coding agent remains the editing surface.

10. Context-quality evals

    • Keep the A/B pattern that measures whether lane context improves agent answers.
    • Each commissioned repo gets its own cases and answer keys.
    • Executable mapping: context/manifest.json.contextQuality commissions the provider-neutral case/answer identities and improvement policy; the matching evals/results.json suite binds the exact manifest and proves paired baseline/candidate uash.context-arm-result.v1 JSON documents through uash.context-comparison.v1. The gate derives each aggregate and critical-regression count from ordered per-case evidence, so detached declared scores cannot pass.

Move into the generated project adapter

These must not live in the universal core:

Adapter field	Why it is project-specific
Product/company names	Different per repo/team
Users/customers	Defines blast radius and UX expectations
Production definition	Branch/deploy models differ wildly
Merge/deploy owner	Human authority differs by org
Source-of-truth order	Some teams trust GitHub first, others Linear/Jira/docs/live dashboards
Validation commands	Every stack has different tests/build/evals
Safe edit paths	Depends on architecture and ownership
Review-required paths	Depends on auth/data/billing/infra surfaces
Red Zone actions	Risk profile differs by product
Answer style	Human preference differs
Lane set	Repos need different work lanes
GitNexus/code-intelligence path	Generated per repo snapshot

Product rule

Universal core = the rules of the operating system.
Project adapter = the local laws of this repo/team.
Agent connector = how Claude Code/Codex/Hermes obey the rules.
Run packet = the evidence that they did.

Practical implementation in this repo

• scripts/commission-harness.mjs is the first local commissioning CLI.
• docs/UNIVERSAL_COMMISSIONING_FLOW.md describes the product flow.
• templates/claude-code/commands/valdris-sdlc-harness.md is the generic Claude Code slash-command template.
• scripts/claude-code-bridge.mjs and scripts/uash-emit-event.mjs implement the local event bridge.