Clearer progress updates and user communication
This patch updates shared writing guidance across Claude agents and core commands, Codex workflow skills, and generated Codex agents. It retains Orwell's six rules and adds guidance shaped by Microsoft's article on natural AI writing and a bounded live check.
Writing guidance
- Match the reader's background and requested detail. Explain unfamiliar terms when needed.
- Use progress updates to report findings, decisions, blockers, and what the next check will resolve.
- Describe observable user behavior and keep hypothetical examples clearly labelled.
- Summarize routine checks; include internal counts and tool details when they affect the reader's next decision.
- Keep uncertainty, failures, and remaining actions explicit. Preserve exact technical details, required report sections, and JSON-only formats.
Documentation
The README and wiki now describe the current agent responsibilities, presentation and debate modes, migration procedure, Ember setup, and recovery options. The agent reference includes illustrative communication examples.
Validation and limits
Two live agents produced four sampled replies covering a handoff, stakeholder JSON summaries, and a hypothetical progress update. The first stakeholder summary exposed excess internal detail; revised guidance reduced that detail in the retest. These samples do not establish consistent behavior across hosts, measured productivity, or user-study results. The release announcement records the final local regression results.
Updating
Existing 4.6.0 installations can use the current updater. For the first upgrade from legacy agent names, use the source-checkout migration procedure with a reviewed release commit. Restart Claude Code and Codex after installation so they load the revised agents and skills. GitHub Actions remains disabled.