Six profiles assign the root, workers, and reviewer
Version 0.2.1 ships 6 profiles. Four use the original Pro or Plus arrangement, with variants capped at 2 or 4 concurrent subagents. Two newer profiles use GPT-6 Sol at medium or max effort as root and reviewer, while GPT-6 Luna handles execution. The Pro default puts GPT-6 Astra at the root, Luna on explorer, worker, tester, and researcher roles, and Astra on review.
The split is concrete rather than decorative. Each role has its own TOML file with a model, reasoning effort, description, sandbox setting, and developer instructions. The worker is told to stay inside a bounded implementation scope. The explorer is read-heavy, the tester validates, and the reviewer reports findings back to the root. One astra-orchestrator skill defines how the root should decide, spawn, wait, integrate, and verify.
What happened when we ran it
Our sandbox installed commit 30b7d0b in 21 seconds, pulling 35 packages and using 37 MB. The build passed in 8 seconds. Pytest completed in 9 seconds with 9 passed and 0 failed. Pip-audit reported 0 known vulnerabilities. The container had 3 CPUs, 8 GB of RAM, Python 3.12, no secrets, and no elevated privileges.
The repository is small: 58 files, about 871 lines of source, and a 0.2 MB checkout. It contains a tests directory, no Dockerfile, and no CI workflow. A green local suite therefore verifies the profile and token-script checks we ran, but GitHub is not automatically repeating that evidence on every push through a checked-in workflow.
The installer asks before replacing project files
The shell and PowerShell installers target a different, existing repository. After you choose 1 of 6 profiles, the script offers .codex, .agents, and AGENTS.md separately. It lists existing paths that would be replaced and defaults update confirmations to no. Existing AGENTS.md content is preserved and the orchestrator instructions are appended if they are not already present.
That is better than blindly copying a preset into a mature codebase. It still changes the rules that Codex reads whenever it works in the target project. Official OpenAI documentation says project .codex/config.toml files load only for trusted projects, and project config outranks user defaults. Review model choices, approval policy, sandbox mode, concurrency, and every custom-agent instruction before confirming the copy.
The global-install instructions need even more care because they tell you to merge selected settings into your personal Codex configuration. A project install is easier to audit, diff, and remove. Start there, then promote only settings that have proved useful across several repositories.
The binary delegation gate is too eager for modest work
Issue 5 reports that simple tasks expanded into explorer, worker, tester, reviewer, and extra-specialist cycles, increasing token use and making work slower. The installed skill at commit 30b7d0b still uses a binary gate: multiple files, external research, cross-component debugging, or useful independent review can make delegation mandatory. An open pull request proposes a 3-tier gate, but that change is not part of this reviewed commit.
Official OpenAI documentation makes the cost direction plain: each subagent performs its own model and tool work, so subagent workflows use more tokens than comparable single-agent runs. Parallel exploration can save wall time when tasks are independent. Parallel writing can create conflicts and coordination overhead. The repository's one-writer-per-file guidance helps, but the first decision remains whether the task needed extra threads at all.
One historical run used 9.3 million total tokens
The token guide includes a single historical cross-component bug fix with 4 spawned roles and a wall time of 13 minutes 49 seconds. It reports about 9.3 million total tokens, of which 96.3% of input tokens were cache hits. The guide correctly tells readers to focus on uncached input, output, and rate-window movement instead of treating the total as equivalent billable work.
That sample predates the current profiles, so it cannot prove that Pro, Plus, or either Sol profile is cheaper or better. It does show where overhead sits: the root stays alive, polls agents, and rereads its context. The bundled read-only token_usage.py groups Codex rollout files by root session and role, giving you a way to compare root-only work with the installed profile on your own tasks.
v0.2.1 is active, but model access is outside the repo
GitHub showed 1,657 stars, 6 open issues and pull requests, and a last push on September 23, 2026. Release v0.2.1 landed the same day and moved the original profiles to GPT-6 Luna while preserving their effort levels and subagent caps. Recent pull requests cover both shell and PowerShell installation paths, and our 9-case suite passed.
The repository cannot grant access to Astra, Sol, or Luna. Official OpenAI guidance says to choose a model available to your signed-in account or workspace, and custom agent files can override both model and reasoning effort. Issue 11 also reports that model selection is visible for new conversations but not existing conversations or branches. Install into a disposable project first, start a new Codex session, inspect every role, and compare one genuinely parallel task with a root-only baseline before standardizing the setup.