openmozi: Instructions file for Codex

AGENTS.md — Repository Instructions for Coding Agents

This file applies to Codex, Claude Code, Gemini CLI, and any other coding agent working in this repository.

The repo-level source of truth is docs/CONSTITUTION.md. Follow it for all architecture, delegation, and release decisions.

Non-Negotiable Rules

1. Prompt text is not the execution engine. Runtime owns worker launch, sandboxing, health, state, fallback, and verification.
2. Never invent worker progress, fake queued/completed states, or hide a deterministic runtime failure behind vague generic apologies.
3. Only claim capabilities that are actually registered and currently ready in this runtime.
4. Treat user-defined workspace skills and workspace agents as first-class extensions; they must follow the same managed-worker contract as built-in flows.
5. Do not ship delegation changes unless a real complex task has been driven end-to-end on the build and the evidence recorded. The automatic gate was removed (see docs/CONSTITUTION.md §14) — nothing checks this for you.
6. Do not use generic shell execution as the transport for Claude Code, Codex, Gemini, or other external AI CLIs; those belong to managed-worker skills/adapters.
7. Do not treat a feature change as complete until relevant local tests have been run and passed; if new behavior lacks coverage, add or update tests in the same change.
8. Treat every bug report or feedback item as investigation-only until the user reviews the evidence and explicitly approves an implementation option. Follow the decision gate in docs/CONSTITUTION.md §6.
9. Load skills/mozi-development/SKILL.md before adding or changing anything in this codebase. It carries the existing seams to reuse and the wiring proof a change needs; a raw code scan misleads you here (dead exports, docs describing intent rather than behaviour, a second implementation winning silently) and unit tests do not catch it.
10. Never publish to the open-source repository freehand. Load and follow skills/public-release/SKILL.md every time — it carries the review gates the export script cannot enforce. Nothing that only makes sense inside this tree may appear in the public one: not in files, not in a commit message, not in a tag annotation, not in a release note. This has already gone wrong once (the tooling's own commit message published this project's name, version and commit sha on every sync) and cost a rewrite of published history.