Only when selected: gpt-6-luna with xhigh. This changes execution settings, not workflow, authority or acceptance.
Use workspace config and repository role in the selected workspace's .codex/config.toml and .codex/agents/. Preserve unrelated settings/credentials; no user-home installation or trust bypass. Existing sessions do not change. Verify actual host/model support; prompts and valid TOML do not prove a model run. Pinned role settings may override spawn choices; choose a compatible role.
The only role is repostew-repository. Discovery, implementation, audit and
review are packet phases with their own scope, not separate role configurations.
Remove retired registrations during migration; cached host roles require reload.
Root owns queue, admission, shared state, jobs and evidence-backed acceptance. Use one repository leaf per repo. Complexity alone never blocks; return actual uncertainty/evidence/recovery, not invented coverage.
Follow inline dispatch: full stable policy first, packet last, fresh leaf for each new repo, owner/repo name, same-repo delta only. When supported use model="gpt-6-luna", reasoning_effort="xhigh", fork_turns="none" as host API arguments, not TOML keys. Acknowledge full-history fallback cost; never invent reset/seed APIs. Record returned identity/actual occupancy.
Read full included policy once; restore missing content after compaction/revision. Keep logs in durable evidence and returns compact. Shared files do not remove per-leaf context cost; identical prefix does not guarantee cache hits or subscription discounts. Measure actual exposed usage; xhigh reasoning still costs tokens.