Fable Super God Mode

Fable Super God Mode

What this adds

This skill adds ONE thing to Fable God Mode: a GPT-5.5 lane, reached through a vendored one-shot bridge to the Codex CLI. GPT-5.5 acts as Fable's deterministic specialist and independent critic on correctness-heavy work.

Fable God Mode's 10-80-10 discipline still governs everything — Fable 5 plans, cheaper Claude subagents execute, Fable reviews and owns final quality. Nothing here replaces or relaxes that loop; see skills/fable-god-mode/SKILL.md (installed as the fable-god-mode skill). This skill only inserts a second, independent model at specific correctness checkpoints inside that same loop.

When to use the GPT-5.5 lane

Keep it gated. Fire the lane for:

• Deterministic / backend / algorithmic work — numeric, boundary, parsing,

state-machine, concurrency, data/scoring math, or API-contract correctness.

• Plan critique — before executing a multi-step deterministic plan.
• Diff review — of substantial deterministic changes before finalizing or

committing.

• Debugging second opinion — AFTER one failed fix attempt, not before.

Do NOT fire it for creative work, frontend / visual / UX, copy / prose / docs, conversational turns, or trivial edits — a second model there adds cost without decisive signal. Full routing detail lives in references/routing.md.

Running a critique

Three steps: write a prompt file, invoke the bridge, read the envelope.

1. Write the prompt file. A plain text/markdown file containing the FULL,

self-contained critique request: the task context plus the material to review inline, or paths for Codex to read. The bridge sends only this file (plus a fixed output-contract suffix) — never your environment. Codex runs read-only-sandboxed and may READ files in the working directory only if the prompt points it at paths.

1. Invoke the bridge (Node ≥18, zero dependencies):

``sh node "${CLAUDE_SKILL_DIR}/scripts/ask-codex.mjs" [--model ] [--timeout ] ``

Model selection: --model flag > CODEX_MODEL env var > default gpt-5.5 (always passed explicitly to the CLI). The CLI is run as codex exec --sandbox read-only --ephemeral, non-interactive.

1. Read the envelope. The bridge prints ONE JSON object on stdout:

``json {"verdict", "model", "summary", "findings": [...], "error", "elapsed_ms"} ``

verdict is one of approved | findings | codex_unavailable. Each findings item is {severity, category, where, issue, suggestion} where severity ∈ critical | high | medium | low.

Exit codes: 0 = approved, 10 = findings, 20 = codex_unavailable, 2 = usage error (bad arguments — a caller bug you must fix, not an outage).

Handling the verdict

Three states, and you must treat them differently.

• approved — GPT-5.5 found nothing. This ADDS to your own review; it does

not replace it. Ship on the union of both.

• findings — adjudicate below.
• codex_unavailable — this is the single most important rule.

Tri-state rule. codex_unavailable means NO review happened. You MUST surface this plainly to the user — "GPT-5.5 review did not run: ``" — and proceed on your own findings only. An outage must NEVER be reported as a clean review, and must NEVER be silently ignored. This is a deliberate improvement over fail-open designs that print "approve" on error.

Union rule. GPT-5.5 findings ADD to your own self-review findings. GPT's approval or silence NEVER drops a finding you already had. Adjudicate each GPT finding on merit: accept the real ones and act on them; reject false positives with a stated reason. The result you ship is your findings ∪ the accepted GPT findings.

Log every critique as one JSON line appended to the project's .orchestration/feedback.jsonl:

{"ts": "", "artifact": "", "reviewer": "gpt-5.5", "verdict": "", "findings": , "accepted": , "rejected": , "notes": ""}

Data disclosure

Using this lane sends the prompt file's contents — and any files Codex reads under it — to OpenAI. The user consented to this at install. Do not send secrets, credentials, or files the user has not approved for disclosure.

Setup

Any paid ChatGPT plan (Plus and up) includes the Codex CLI; the free tier has limited quota. The full install / login / verification procedure lives in references/setup-codex.md — defer to it.

Advanced path: AgentBridge (github.com/quilin-ai/agent-bridge, MIT) enables live bidirectional Fable↔Codex sessions; see references/routing.md for when that is worth it. It is NOT installed or configured by this skill's installer.

References

Load these on demand:

• references/routing.md — full routing detail: which lane fires when, the

NOT-for list, and the AgentBridge advanced path.

• references/setup-codex.md — Codex CLI install, login, and verification.
• references/verdict-schema.json — the JSON Schema enforced on GPT-5.5's raw

verdict (approved/findings only; codex_unavailable is the bridge's own state, added in the envelope along with model, error, and elapsed_ms).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: nagarjuna-msr
• Source: nagarjuna-msr/fable-god-mode
• License: MIT

Install and usage instructions live in the source repository linked above.