Decide whether the work needs separate agents
Codex Astra Luna Orchestrator supplies a ready-made division of work. The decision is whether your tasks benefit from that division enough to justify additional agent activity.
Issue 5 describes a user's experience on a production Laravel codebase: relatively simple tasks expanded into explorer, worker, tester/reviewer, and orchestrator cycles. The reporter says this increased context use and occasionally slowed straightforward work. This is a user report, not our benchmark or proof that every profile behaves that way.
Official Codex documentation explains the tradeoff: each subagent performs its own model and tool work, so comparable subagent workflows consume more tokens than single-agent runs. Independent exploration can benefit from parallel work; simultaneous edits can create conflicts and coordination overhead. Use that distinction when deciding which assignments deserve separate agents.
Six profiles, with a dated model distinction
The README at reviewed commit 30b7d0b lists six profiles. Pro uses GPT-6 Astra at medium effort as root, GPT-6 Luna at max effort for execution, and Astra at low effort for review. Plus uses Luna at max effort as root, Luna at medium effort for execution, and Astra at low effort for review. Both have four-subagent and two-subagent variants.
The two Sol profiles at that commit use GPT-6 Sol at max or medium effort for root and reviewer, with Luna at max effort for execution and a four-subagent cap. The supplied current README instead specifies GPT-6.1 Sol for the Sol orchestrator and reviewer. The profile directory names remain GPT6-SolMax-LunaMax and GPT6-SolMedium-LunaMax. Our recorded run does not validate that later configuration.
Release v0.2.1 describes the original four Pro/Plus profiles moving their Luna pins from GPT-5.6 to GPT-6 while retaining their reasoning levels and subagent limits. Its statement that all nine tests passed is the maintainer's release verification, separate from our stored run.
What our recorded run establishes
Our run is dated September 29, 2026, at commit 30b7d0bb7e9b3b9a7d27e78a11c468f95c4cf704. The stored sandbox configuration specifies 3 CPUs, 8 GB of memory, and image ghcr.io/astral-sh/uv:python3.12-bookworm. The install tail identifies Python 3.12.12.
| Recorded stage | Observation | Limit |
|---|---|---|
| Install | Exit 0 in 21.428 seconds; dependency summary records 35 packages and 37 MB | The available tail only identifies the Python environment; it does not establish a Codex or profile installation |
| Build-labelled step | Exit 0 in 8.245 seconds | Its tail is empty, so the record does not establish what was built |
| Test | Exit 0 in 9.371 seconds; output reports 9 passed and 23 subtests passed in 0.61 seconds | Parser summaries do not establish exhaustive unique coverage |
| Audit | pip-audit records zero advisories | Coverage is unknown; zero advisories is not proof of safety |
The stored harness total is 55 seconds. The rounded install, build, and test durations sum to 38 seconds, but that sum is neither installation time nor the complete run duration.
The checkout summary records 58 files, approximately 871 lines of source, 0.2 MB, a tests directory, no detected CI workflows, and no Dockerfile. These are observations about that checkout. We did not run a delegated coding workload, verify installed-binary behavior, measure model access, or compare cost and task quality.
Review the project changes before installing
The README directs users to clone this repository, run setup.sh or setup.ps1, select an existing target repository other than the installer repository, and choose a profile. It says the installer copies configuration into .codex/, the skill into .agents/, and project instructions into AGENTS.md. Existing component files require confirmation before updates; existing AGENTS.md content is preserved, with instructions appended when that file exists.
Codex loads project configuration only for trusted projects. Project configuration takes precedence over user defaults, while higher-priority overrides and managed requirements still apply. Inspect the selected model settings, concurrency, project permissions, and instructions before accepting those changes.
Custom agent instructions do not create an independent security boundary. Current official documentation says custom agents inherit the parent's permissions, provider, and MCP servers; placing sandbox or approval settings in a custom agent file does not override the parent configuration.
The README also describes global installation and tells users to merge settings rather than overwrite existing Codex configuration. A project-scoped trial gives you a smaller set of changes to inspect first.
Confirm the models in a new session
Profile names do not establish account entitlements. Confirm that your client and account can use every configured model. Issue 11 reports that initial model selection appeared in new conversations but not existing conversations or branches. Treat this as a reported limitation, not a universal compatibility result.
Install into a disposable target repository, inspect the resulting files, start a new Codex session, and check the selected roles. Then compare an independently divisible task with your existing workflow. Record task completion, review findings, elapsed time, and token use separately. Our run supplies no savings benchmark for either the reviewed or current profiles.
How the ratings work
These are editorial ratings of the supplied evidence, not measured performance scores. On the five-point scale, 1 means substantial obstacles or very little support, 2 means partial support with major unknowns, 3 means a documented basis with meaningful validation limits, 4 means strong support with bounded gaps, and 5 requires broad corroboration relevant to the category. Missing adoption, compatibility, or workload evidence earns no positive credit.
Setup receives 3/5 because the documented installation path includes confirmation and preservation steps, but our run does not validate interactive installation or model access. Documentation receives 3/5 because the README explains profile assignments and installed paths, while live compatibility remains untested. Community receives 2/5 because the supplied issues contain concrete user feedback but do not establish adoption or support responsiveness. Maturity receives 2/5 because v0.2.1 and local test output provide some validation, while the reviewed checkout has no detected CI workflow and no measured orchestration workload.

