Agent Ops

agent-ops is the operational reality layer for coding agents.

It should answer the questions code search cannot:

• Did the branch pass CI?
• Is the latest commit actually deployed, and where?
• Did runtime errors appear after that deploy?
• Is the PR review-ready or release-blocked?
• Who owns the problem if the system is unhealthy?

Honest current state

Today this repo is a read-only MCP server with:

• GitHub workflow status for the current branch
• environment-aware deploy health checks with per-env commit comparison
• a Sentry unresolved-issues snapshot
• PR review and comment snapshotting
• a normalized get_ops_summary() with full ops-context schema
• explicit release-readiness computation with environment-aware blocker logic
• append-only ops history with fingerprint dedup
• a memory bridge to Agent Memory for durable ops findings
• local config storage in .agent-ops/config.json

Still missing:

• incident provider integrations and regression detection from recurrence evidence
• ownership/escalation signals
• richer deploy/runtime providers beyond a generic health endpoint
• a standalone operator-friendly CLI surface

Why this repo exists

The rest of the toolkit already covers adjacent truths:

• agent-memory preserves execution continuity
• agent-coordinator shows who is doing what now
• agent-guardian validates change safety
• agent-project-context explains business and product intent

agent-ops should tell an agent what is true in CI, deploys, runtime, and release state right now.

When an agent should consult agent-ops

• after pushing code
• before merging or handing off work
• at session boot for build, deploy, release, or incident tasks
• when a human asks "did this actually ship?" or "is prod healthy?"
• when deciding whether a failure is local, pipeline, deploy, or runtime

Current MCP tools

• check_ci() - branch-level GitHub Actions status
• check_deploy() - deploy health plus commit comparison (legacy single-environment)
• get_deploy_matrix() - per-environment deploy state across all configured environments
• check_errors(since?) - Sentry unresolved issue snapshot
• check_pr() - PR state, reviews, and comments for the current branch
• get_ops_summary() - full normalized operational context in one call
• get_release_readiness() - release readiness with blockers and evidence
• record_ops_snapshot(nodeId?) - persist snapshot to history and optionally bridge to Agent Memory
• configure_ops(config) - save integration config with env-var token refs

Read this repo in this order

1. README.md
2. ROADMAP.md
3. PAIN-POINTS.md
4. docs/IDEAL-OPS.md
5. docs/OPS-CONTEXT-SPEC.md
6. CLAUDE.md

Local setup

1. npm install
2. npm run build
3. powershell -ExecutionPolicy Bypass -File scripts/bootstrap-memory.ps1
4. Open this repo in Claude Code or Codex with the local .mcp.json

Agent Memory

This project uses local Agent Memory via .mcp.json.

If boot_context() does not recover a usable active node, run scripts/bootstrap-memory.ps1 before planning or editing. The bootstrap script seeds the repo with a project node, a completed baseline feature, and the current active feature/task for the next implementation slice.