Self-Development Meta-Workflow Specification

Version: 2.4.0 Status: COMPLETE (All Implementation Phases Done) Date: 2026-01-09 (ET) GitHub Issue: #144

Prerequisites

Prerequisite	Description	Tracking
v2.3.0 CLI Adapters Stable	Claude, Gemini, Codex adapters production-ready (1 day 0 critical bugs)	Issues #75-#77
v2.3.0 Quality Router	Quality-based model routing tested	Issue #78
Docker Available	Docker CLI available for sandboxed execution	Verified
Security Audit Complete	All security safeguards implemented and tested	This document

Stability Definition: Zero critical bugs for 1 calendar day after deployment.

Executive Summary

This specification defines a meta-workflow for nexus-agents self-development. The workflow enables the system to analyze open issues, plan implementations using its own protocols (TRINITY, Consensus, Reflexion), achieve multi-agent consensus, obtain human approval, and execute implementation using Self-Debug and Self-Refine protocols.

Security Model: All code execution occurs in Docker containers with strict resource limits. Human approval is required for plan approval only (Phase 6). After approval, automated security gates (input sanitization, Docker sandbox, security scans) handle execution. PRs are created for milestone changes to enable tracking and rollback. Only repository owner can trigger workflows.

Autonomy Model: After human approves the plan, the workflow runs autonomously through security gates. Humans receive notifications but are not blockers. Rate limiting is optional and disabled by default.

Workflow Overview

 +-------------+     +-------------+     +-------------+
 |   ANALYZE   |---->|  RESEARCH   |---->|    PLAN     |
 | (Sanitized) |     | (Context)   |     | (TRINITY)   |
 +-------------+     +-------------+     +-------------+
        |                  |                   |
        v                  v                   v
 +-------------+     +-------------+     +-------------+
 |    VOTE     |<----|   REVIEW    |<----|   REFINE    |
 | (Consensus) |     | (Human)     |     | (Reflexion) |
 +-------------+     +-------------+     +-------------+
        |
        | [PLAN APPROVED - Autonomous from here]
        v
 +-------------+     +-------------+     +-------------+
 |  GENERATE   |---->|  IMPLEMENT  |---->|   SECURE    |
 | (Code)      |     | (Sandboxed) |     |   CHECK     |
 +-------------+     +-------------+     +-------------+
        |                  |                   |
        v                  v                   v
 +-------------+     +-------------+     +-------------+
 |   VERIFY    |---->|   COMMIT    |---->| MILESTONE   |
 | (Tests)     |     | (PR)        |     |   PR        |
 +-------------+     +-------------+     +-------------+
                                               |
                                      [Notify human, auto-merge]

Phases Summary

Phase	Name	Protocol(s)	Output	Human Checkpoint
1	Analyze	Adaptive + Input Sanitization	Prioritized issue list	No
2	Research	Parallel execution	Context & prior art	No
3	Plan	TRINITY (Thinker/Worker/Verifier)	Implementation plan	No
4	Refine	Reflexion (multi-persona critics)	Refined plan	No
5	Vote	Consensus (5-agent vote)	Approval/rejection	No
6	Review	Human checkpoint	Plan approval	YES
6.5	Code Review	Human checkpoint	Code approval	YES
7	Implement	Self-Debug (Docker sandbox)	Working code	No (automated)
7.5	Security Check	SecureCodeChecker	Security scan results	No (automated)
8	Verify	Test execution	Test results	No (automated)
9	Commit	Git operations	Milestone PR	No (notify only)

Note: After Phase 6 (Plan Approval), the workflow runs autonomously. Security is enforced by automated gates (Docker sandbox, security scans, test verification).

Sub-Documents

This workflow specification is split into focused sub-documents for readability:

Document	Contents
Phases 1-6	Analyze, Research, Plan (TRINITY), Refine (Reflexion), Vote (Consensus), Review (Human)
Execution Phases	Code Review, Implement (Docker sandbox), Security Check, Verify, Commit
Operations	Error handling, configuration, metrics, rate limiting, rollback, runbook, audit trail
Validation & Implementation	Improvement validation protocol (Tier 1-3 gates), WIS framework, implementation notes

Research Integration

Paper	Technique	Integration Point
arXiv:2512.04695	TRINITY	Phase 3 (Plan)
arXiv:2512.20845	Multi-Agent Reflexion	Phase 4 (Refine)
arXiv:2303.17651	Self-Refine	Phase 7 (Implement)
arXiv:2304.05128	Self-Debug	Phase 7 (Implement)
arXiv:2502.19130	Adaptive Selection	Phase 1 (Analyze)

Approval

This specification requires multi-agent consensus before implementation:

Agent	Vote	Notes
Architect	APPROVE	v2.0.0 security model
Security	APPROVE	Docker sandbox adequate
DevEx	APPROVE	Human checkpoints clear
AI/ML	APPROVE	Protocol selection sound
PM	APPROVE	Risk/value balanced

Threshold: Supermajority (4/5) required. Result: Unanimous (5/5).

Specification created: 2026-01-08 (ET) Revision: 2.4.0 - Split into sub-documents for governance compliance Protocol versions: TRINITY v1, Reflexion v1, Consensus v1