flnzba.github.io: Instructions file for Codex

Engineering Agent Instructions

Applies to every agent session here and in template-derived repositories.

Before deciding

• Inspect current code, tree, and effective runtime/config. Stale or unverified artifacts cannot override observed state or user-approved requirements.
• Query the CodeGraph MCP index first for structure, symbols, callers, and impact; explore by hand only what the index does not answer. The index persists across sessions — do not re-derive what it already knows.
• Use the software-engineering skill for every substantive task: requirements, architecture, implementation, debugging, or review. It loads automatically; never wait for a slash command.
• Derive before you ask. The repo, git history, the user's own sites, package registries, vendor docs and installed CLIs answer most factual questions. Before calling an external dependency blocking, check whether its CLI is authenticated and whether the account already holds what you need.

While working

• Prefer the smallest correct change. Reuse native and existing capabilities before adding code, configuration, or dependencies. A datastore nothing queries, a service nothing calls, an endpoint nothing hits: build each when something needs it, so it never becomes the scope you delete later.
• Keep requirements and product decisions in the main conversation. Ask only questions whose answers change what you will build.
• Resolve root causes. Never suppress, work around, or defer an error to keep moving.
• Edit canonical sources; regenerate and verify derived outputs. Never hand-edit copies or assume config is effective.
• Re-scan shared state before mutations. Never revert, overwrite, or clean up work you did not make. Keep secrets out of commits, logs, and tool output.
• Follow the repo-wide service architecture standard (see the software-engineering skill's architecture reference): DTOs at API boundaries, data access behind DAOs/repositories, services communicate only through published APIs. Use repository-managed runtimes and commands; Python runs only through the project venv or a managed runner. Keep hermetic checks independent of ambient credentials and network.
• Configuration is centralized (see the skill's platform reference): committed non-secret env bases plus one gitignored secrets file whose keys are all listed in a committed example; no per-service env files, no environment literals in code.
• Infrastructure keeps three tiers (platform reference): leaf modules → one root stack → thin per-environment roots. Topology is declared data, secrets are injected, and generated config carries a drift check.
• Delegate independent read-heavy work (exploration, research, review) to subagents; never run concurrent editing agents. Independently verify their claims. The current user request outranks stale hook or TODO work.