Skip to content

AGENTS.md

Latest commit

 

History

History
34 lines (22 loc) · 2.95 KB

AGENTS.md

File metadata and controls

34 lines (22 loc) · 2.95 KB
Raw

Agent guide (Codex / OpenCode / any non-Claude runtime)

Read CLAUDE.md — it is the canonical agent guide for this workspace. Everything there applies to you regardless of runtime: the tool inventory, the Database safety contract, the RCA workflow, and the agent guidelines. This file only covers what differs outside Claude Code.

Start here: preflight

Run ./scripts/reckon preflight before the first query of any task. It prints the active environment, which integrations are genuinely usable, which are not and why, what knowledge exists, and any warnings — without touching infrastructure. Trust it over the toolbelt list in CLAUDE.md: that describes what reckon can connect to, preflight describes what is connected. ./scripts/reckon doctor diagnoses setup problems; ./scripts/reckon use <env> switches environment.

direnv is optional — if it isn't hooked into your shell, .envrc never loads and nothing warns you. Activate manually with eval "$(./scripts/reckon env)".

Investigation methodology

The full RCA methodology lives at skills/reckon/SKILL.md (with references in skills/reckon/references/). Claude Code loads it as a skill automatically; if your runtime has no skill loader, read that file at the start of any incident investigation and follow it. Consult infra-knowledge/ before querying anything.

Things that don't transfer from Claude Code

  • Permission prompts — Claude Code's allowlist enforces the "every DB query prompts for approval" contract. On other runtimes, use the equivalent (Codex: approval mode untrusted/on-request; never run with full-auto against this workspace), because the DB clients touch production.
  • Skills — the grafana/jenkins/cubeapm command-reference skills under .agents/skills/ are plain markdown; read them like documentation.

Billing / keys

Subscription logins (Claude Pro/Max via claude → /login, ChatGPT via codex login) need no keys in this workspace. API-key billing reads ANTHROPIC_API_KEY / OPENAI_API_KEY from .env (loaded by direnv). scripts/agent.sh launches whichever runtime is available or requested.

Hard rules (same as CLAUDE.md, restated because they are load-bearing)

  • This workspace is production-only and read-only by usage: never write to a DB, never kubectl apply/delete/scale, never trigger a Jenkins build, never produce to Kafka — and never consume with a group id (kcat -G / rpk topic consume -g), which joins and rebalances a production consumer group.
  • Re-read the Database safety contract in CLAUDE.md before any psql/mysql/mongosh query. EXPLAIN first, LIMIT always, and invoke mysql via mysql --defaults-extra-file="$XDG_CONFIG_HOME/mysql/my.cnf".
  • Write all RCA output to incidents/<YYYY-MM-DD>-<slug>/ per skills/reckon/references/incidents-convention.md — never to the repo root.