Autonomous Healthcheck Workflow for Frontier Model Agents

Objective

Create a .agent/workflows/healthcheck/ protocol that frontier model agents (Antigravity, Gemini CLI, Claude Code) execute periodically (daily/weekly) to diagnose and treat system-level degradation across the Agent OS infrastructure.

Rationale

The Neo.mjs Agent OS currently has no mechanism for detecting "silent degradation" — scenarios where the system appears functional but delivers fundamentally wrong context (e.g., stale session summaries, unreachable ChromaDB, corrupt graph topology). Historical evidence shows that every infrastructure recovery required human pattern recognition and manual intervention (the "doctor visit" pattern observed across 10 sessions in 30 hours, April 12-14 2026).

This workflow bridges the gap between "healing sessions" (human-directed) and "self-healing" (autonomous). A frontier model has the reasoning capacity to correlate cross-system anomalies, write targeted Playwright tests, fix source code, and verify the fix — all capabilities an SLM cannot reliably perform.

Design

Scope

This is a .agent/workflows/ artifact (not a Neo agent profile or npm run script). It is a Markdown protocol that a frontier model reads and follows, similar to .agent/skills/.

Protocol Phases

Phase 1: Infrastructure Pulse

• healthcheck() all 4 MCP servers (KB, Memory Core, Neural Link, GitHub Workflow)
• Verify ChromaDB collection counts and accessibility
• Check SQLite graph node/edge counts + last modification timestamp
• Verify DreamService last successful run

Phase 2: Pipeline Freshness

• Read sandman_handoff.md — is the data current?
• Check summarization backlog (total sessions vs. summarized count)
• Verify latest session summary timestamp vs. current time (staleness detection)
• Confirm Golden Path recommendations reference open tickets (not closed/stale)

Phase 3: Data Quality Audit

• Sample 5 recent session summaries — do quality scores and content make sense?
• Query graph for orphaned nodes (nodes with 0 edges)
• Check embedding model consistency across collections
• Validate that get_context_frontier returns relevant, recent context

Phase 4: Treatment (if issues detected)

• For code-level bugs: write a failing Playwright test (per unit-test skill), fix the source, verify the test passes, open PR
• For data-level issues: execute surgical cleanup scripts via services.mjs SDK
• For issues beyond autonomous resolution: create a GitHub issue with full diagnostic context and signal BLOCKED state

Phase 5: Report

• Generate a structured health report in sandman_handoff.md or a dedicated artifact
• Flag anomalies with severity (CRITICAL / WARNING / INFO)
• Save diagnostic session to Memory Core for future reference

Key Constraint

This workflow MUST be executable by a frontier model agent (Antigravity, Gemini CLI, Claude Code) using their native tool access. It must NOT depend on npm run or Neo agent infrastructure, since those components may themselves be the failure point.

Acceptance Criteria

• .agent/workflows/healthcheck/WORKFLOW.md protocol document exists
• Protocol covers all 5 phases with explicit tool invocation sequences
• At least one real diagnostic scenario is documented with expected tool calls
• Workflow is referenced in AGENTS_STARTUP.md for discoverability

Origin Session ID: 69b726ca-97c8-406c-9106-c63b6642c4c0