- v1.0.0 released 3 Sep 2026
- Every model pin proven from its session record
- 143 offline tests, zero tokens spent
Risk-gated multi-model orchestration for Claude Code and Codex. Claude Fable writes the specification and accepts the result; pinned GPT-5.6, DeepSeek and GLM lanes perform it; a fresh read-only reader from a different model family returns one verdict — ship, fix-first or rethink. Every lane's model, effort and sandbox are proven from the runtime's own session record, and a mismatch refuses the report.
Before any task tool is called, the wright posts a call sheet — route, cast, reader, independence and risk — in a form a machine can audit. There are five routes, cheapest first: solo, delegate, audit, full and ensemble. Solo is the default, because the arithmetic says so: agents spend about four times the tokens of chat, and multi-agent systems about fifteen. The README spells out when not to use the tool at all.
The cast is six pinned lane profiles: luna and terra on GPT-5.6, flash on DeepSeek, glm on GLM, and two reviewers. After a lane runs, the scripts read the runtime's own session record back and check which model, effort, sandbox and provider actually ran. Effort is compared on every turn, so a lane that quietly dropped its reasoning budget halfway through is caught. A routing mismatch exits 69; routing that cannot be verified exits 75. Either way, the report is refused.
Independence is a lookup, not a promise. The reader comes from a different model family in a fixed table — Claude, GPT, DeepSeek, GLM — and a same-family reader has to be declared on the call sheet as residual risk. On Claude Code the reader is a subagent restricted to Read, Grep and Glob, which makes it read-only by construction.
It is POSIX shell and jq, shipped as a plugin for both hosts with a marketplace manifest. The installer refuses to overwrite and rolls back a partial install. 143 offline test cases and a 154-check verify script run on a machine with neither CLI installed, and spend no tokens. v1.0.0 shipped on 3 September 2026.
FABLEWRIGHT and ORRERY are siblings: one lane vocabulary, and one rule that no lane is trusted for work it only claims to have done. The route-gating discipline follows sol-advisor, credited in the README.