Fable review

Fable is the expensive lane, for the calls that are genuinely hard: which of two architectures survives contact, whether a plan's premise holds, whether an interface reads the way its author thinks it does. Use it sparingly and on questions where a second opinion changes what you do next.

What comes back is a set of leads. Verify each one before acting on it.

glm-review is the other lane, and the two are chosen by the kind of question rather than by importance. A bounded conformance check - does this match the stated criteria, does this contract hold, is this copy right - goes to GLM and costs almost nothing. A judgement call with no checkable answer goes here.

One route, three hosts

The Claude CLI is the route, and it works the same from Claude Code, Codex and OpenCode, because all three can run a shell command.