fable-5

Fable 5's working loop for hard, multi-step tasks, packaged as a Claude Code plugin.

Built by LearNer.

Core loop: decompose along verification boundaries → attack the load-bearing unknown first → verify with the smallest check that would fail if you're wrong → pick the next action by what would change the plan.

Contents

• skills/fable-5/ — the skill: 8-step loop, verification catalog (smallest failing check per task type), decomposition patterns, preflight.sh pre-completion sweep.
• skills/fable-5/agents/ — fable-scout (read-only comprehension pass) and fable-refuter (adversarial verifier, defaults to REFUTED when uncertain). Registered automatically on plugin install; also usable as plain prompt templates in other harnesses.

Install

Claude Code (skill + agents auto-registered):

claude plugin marketplace add https://github.com/Learn57130/fable-5
claude plugin install fable-5@fable-5

Codex / Cursor / Kimi — the repo ships .codex-plugin/, .cursor-plugin/, .kimi-plugin/ manifests pointing at skills/; install via each tool's plugin mechanism, or just symlink:

git clone https://github.com/Learn57130/fable-5
ln -s "$(pwd)/fable-5/skills/fable-5" ~/.codex/skills/fable-5

Gemini CLI — install as an extension (gemini extensions install https://github.com/Learn57130/fable-5); gemini-extension.json loads AGENTS.md (the compact loop) as always-on context. Or drop skills/fable-5 into ~/.agents/skills/, which Gemini auto-scans.

Antigravity — no plugin format; clone and symlink the skill:

ln -s "$(pwd)/fable-5/skills/fable-5" ~/.gemini/antigravity/skills/fable-5

Note: if you already expose skills/fable-5 as a personal skill (e.g. via ~/.claude/skills), installing the plugin duplicates it in Claude Code — use one or the other per machine.

Benchmark

Three arms — Opus 4.8 vanilla, Opus 4.8 + fable-5 loop as system prompt, Fable 5 vanilla — graded by held-out deterministic checks with symptom-fix traps (protocol). Arms are isolated from user-level hooks (disableAllHooks). Run 2026-07-07, tasks 04–07 × 3 reps per arm.

xychart-beta
    title "Pass rate on trap tasks, 12 runs per arm (%)"
    x-axis ["Opus 4.8", "Opus 4.8 + fable-5", "Fable 5"]
    y-axis "pass %" 0 --> 100
    bar [100, 100, 100]

xychart-beta
    title "Avg wall time per task (seconds)"
    x-axis ["Opus 4.8", "Opus 4.8 + fable-5", "Fable 5"]
    y-axis "seconds" 0 --> 60
    bar [49.5, 54.2, 56.9]

Honest read. The trap tasks were built so that fixing only the reported symptom fails (two-hop root causes, sibling-caller contracts, hidden docstring requirements, held-out spec edges) — and every arm still passed everything, with clean (hook-free) contexts. Conclusions this data supports:

• Opus 4.8 + fable-5 ≈ Fable 5 on these tasks: indistinguishable (both 12/12).
• Opus 4.8 + fable-5 > Opus 4.8 is NOT shown: vanilla Opus 4.8 also went 12/12 — at this task scale it already root-causes and reads contracts without help.
• The loop costs ~5s/task of extra deliberation and causes no regressions.

Where the loop plausibly pays off is what a cheap headless benchmark can't capture: long multi-step sessions, large codebases, ambiguous scope — places where drift and premature "done" happen. Separating 4.8-class models needs long-horizon tasks with much higher per-run cost. Task contributions welcome: files/ + prompt.md + held-out check.sh.