AgentOps (boshu2)

Repo-native operational layer for AI coding agents. Not a framework — a context-compiler and coordination CLI. Cross-vendor: works with Claude Code, Codex, C…

AgentOps (boshu2)

Repo-native operational layer for AI coding agents. Not a framework — a context-compiler and coordination CLI. Cross-vendor: works with Claude Code, Codex, Cursor, OpenCode simultaneously on the same repo.

GitHub: https://github.com/boshu2/agentops

---

Core Concept

“The moat is the understanding and context you/your team/company compounds. My plugin helps create that moat so it starts compounding and is plugin-proof.”

AgentOps positions itself as vendor-agnostic infrastructure for the accumulated context that makes agents productive. As individual agent tools (Claude Code, Codex, etc.) get absorbed or change, the durable corpus stays.

---

.agents/ Directory

The central primitive. A gitignored directory at the repo root:

.agents/
  decisions/      # architecture decisions, ADRs
  constraints/    # known limits, tech debt notes
  patterns/       # discovered code patterns
  sessions/       # per-session summaries
  council/        # multi-vendor consensus runs

Gitignored: not committed, not shared with teammates (unless you choose to). Survives session clears. Not a second CLAUDE.md — it’s an operational scratchpad that agents read at session start.

Why gitignored: reduces doc-rot risk. Documents here are meant to be continuously rewritten as understanding evolves, not committed as authoritative history.

Contrast with Pocock workflow: Pocock commits PRD/issue files and deletes them after ship (closed → purged). AgentOps keeps a durable gitignored corpus. Neither is strictly better — tradeoff between git history visibility and doc-rot.

---

/council — Multi-Vendor Consensus

The standout feature. /council runs a design decision or code review through multiple vendors simultaneously and synthesizes disagreements:

/council "Should we use optimistic locking or pessimistic locking for this transaction?"

→ Claude Code: [analysis + recommendation]
→ Codex: [analysis + recommendation]
→ Cursor: [analysis + recommendation]

→ AgentOps: "Models agree on X. Disagreement on Y — Claude prefers A, Codex prefers B because..."

Disagreements surface as explicit artifacts, not averaged away. This is structured Multi-Vendor Adversarial Review — the /council formalizes it as a workflow command.

---

ao CLI

The control plane:

Command	Description
ao init	Initialize .agents/ corpus in repo
ao context	Compile current corpus into agent-ready context
/pre-mortem	Force agents to list what could go wrong before implementing
/vibe	Generate a quick project status summary
/council	Multi-vendor consensus on a question

---

Relation to Existing Wiki

• Multi-Vendor Adversarial Review — /council is the primary implementation of this pattern
• Dangeresque — also uses cross-vendor review (adversarial reviewer) but at the per-task level
• Domain Glossary (CONTEXT.md) — .agents/decisions/ is a structured form of the shared glossary concept
• Agent Context Instructions — .agents/ corpus serves as extended context instructions
• Lean Agentic Coding Workflow — alternative approach: commit-and-delete vs gitignored durable corpus