Sol Advisor routes Codex work between Sol, Luna and Terra, with a fresh review when risk demands it

The problem Sol Advisor solves, and who it is actually for

Multi-agent delivery has a default failure mode: work gets spread across helpers because the tooling makes spreading easy, and nobody records why. Sol Advisor attacks that by making the route a first-class artefact. The README states that Sol / High "runs the show", declares a risk-gated route before task tools, keeps solo as the default, and uses a single auxiliary only when that improves delivery. The audience is narrow and worth stating plainly: you need a current Codex CLI or ChatGPT desktop app with plugins enabled, GPT-5.6 Sol / High for the primary session, native custom-agent support, and jq. Luna / Max or Terra / High access is only needed when the selected route delegates. If you are not already working inside Codex, this project has nothing for you, and the README does not pretend otherwise.

How the routing mechanism works: one primary, four modes, no silent downgrades

The architecture puts architecture, decomposition, route selection, parent verification, escalation decisions and acceptance in the primary task. Auxiliary work substitutes for root work rather than duplicating it, and the root inspects the complete diff and reruns the requested checks. Four modes are documented. In solo, the default when risk is contained, the root plans, implements, tests and self-reviews. In delegate, a complete spec goes to one implementer: Luna / Max for bounded work, Terra / High for judgment-heavy or high-risk work, with the root verifying. In audit, the root implements and a fresh read-only Sol / High reviews. In full, an explicit broad or high-risk exception, you get one selected implementer, root verification and a fresh Sol / High review. Sol emits a SELECTIVE ROUTE declaration containing the mode and a concise risk rationale before the first task tool call. Escalation is allowed only when newly observed risk justifies it, and the README says the route is never silently downgraded. That last constraint is the interesting design choice: the cost of a wrong route is paid in review, not in a quiet switch.

Installing the Sol Advisor plugin and running the first orchestration prompt

The quick start assumes Codex CLI or the ChatGPT desktop app with plugins enabled, plus jq. The first command adds the marketplace from the main branch, and the second installs the plugin itself.

codex plugin marketplace add DannyMac180/sol-advisor --ref main
codex plugin add sol-advisor@sol-advisor

The third step resolves the installed plugin path from the JSON listing and runs the companion installer, which the README describes as verifying all three exact role files after installation.

plugin_dir="$(codex plugin list --json | jq -r '.installed[] | select(.pluginId == "sol-advisor@sol-advisor") | .source.path')" && test -n "$plugin_dir" && test "$plugin_dir" != null && test -d "$plugin_dir" && test -f "$plugin_dir/scripts/install-agents.sh" && sh "$plugin_dir/scripts/install-agents.sh"

The installer is fail-closed. Modified, unsafe, nonregular, symlinked, unknown or differing files are left untouched, and it does not edit Codex configuration. The README then tells you to start a fresh Codex task so native roles are discovered, and to use this prompt in that new task:

Use $sol-advisor:orchestration to build this feature and verify it. Declare the selective route before task tools.

What you should see is a SELECTIVE ROUTE declaration naming the mode and its risk rationale before any task tool runs. You do not pick a lane yourself; you give Sol the outcome, the constraints and any repository context, and Sol records the route and owns verification and acceptance.

The fail-closed installer is also the sharpest limitation

An installer that refuses to touch anything it does not recognise is a defensible security posture and an operational annoyance. If a role file has been edited locally, or ends up symlinked, or differs from the expected content, the installer leaves it alone. That means a partially installed setup can persist without an obvious error, and the README does not document rollback or a repair path for that state. The bigger constraint is the dependency chain. The primary session needs GPT-5.6 Sol / High, and delegation needs GPT-5.6 Luna / Max or Terra / High. Without those model tiers, the routing table is mostly theoretical. The README also does not state how the workflow behaves if a delegated implementer fails mid-task, how long a route declaration stays valid across a long session, or what happens when the primary context is compacted. Codex-only is a deliberate boundary, but it is a boundary: there is no documented path for running the same routing discipline in another editor or against another provider.

Compared with a plain Codex custom-agent setup

The obvious alternative is Codex's own native custom-agent support, which the README lists as a prerequisite rather than a competitor. Stock Codex gives you agents and lets you decide, task by task, whether to spawn one. Sol Advisor adds a governance layer on top: a named route, a written risk rationale emitted before the first task tool call, a default of solo, a default maximum of one auxiliary, and an explicit exception for full. The difference in approach is where the decision lives. With plain custom agents the decision is implicit in what you type. With Sol Advisor the decision is an artefact you can read in the transcript, and the README states that any fix returned by a review requires a new review. That is a stricter loop than most people run by hand. It is also more ceremony, and if your tasks are small and low risk, the declaration step is pure overhead.

Updating the plugin, and what the MIT licence does and does not cover

Updating is three commands: upgrade the marketplace plugin, reinstall the companion roles, then start a new task.

codex plugin marketplace upgrade sol-advisor
codex plugin add sol-advisor@sol-advisor

After that, rerun the same path-resolution and installer command shown in the quick start. The reinstall step is not optional, because the installer is what verifies the three role files. The repository is licensed under MIT, and the top-level entries are .agents/, .gitignore, LICENSE, README.md and plugins/. MIT is permissive: it lets you use, modify and redistribute the code with the licence and copyright notice preserved, and it comes with no warranty. It says nothing about the model tiers you route to, and nothing about the Codex plugin marketplace terms you accept when you add the marketplace. Whether the licence text covers the role files under .agents/ the same way it covers the scripts is a question for whoever owns your compliance review; the README does not address it. There are no retrieved releases, so upgrades track the main branch rather than a tagged version, and the README does not describe a version pinning scheme beyond the --ref main flag used at install time.

Editorial conclusion

Adopt Sol Advisor if your delivery already happens inside Codex and you want the primary Sol / High session to own routing, verification and acceptance, with delegation treated as an exception that has to be justified. Skip it if you are not on Codex, if you cannot guarantee GPT-5.6 Sol / High for the primary session, or if you want an orchestration layer that works across editors and providers. Before you commit, read plugins/sol-advisor/skills/orchestration/references/operations.md, because the README points there for the exact spawn, runtime-evidence, sandbox and installer rules that the quick start does not spell out.

Frequently asked questions

Official sources

1. DannyMac180/sol-advisor on GitHub
2. Issues
3. License: MIT
4. README