Overview

Before installing and using it for the first time, say "buddy, config": the buddy will open the local single-page UI that only listens to 127.0.0.1. The three starting points—balancing/quality/cost—the CLI/model/effort for the three identities—are all completed on one page; The page first displays the exact diff, and only after confirmation is the data is placed. The beginner process always uses the current project, local Git ignores, and automatic checks after installation—security defaults are used, and advanced options are not pursued. The effort supported by the Codex model and each model comes from the local CLI model/list, Claude alias and effort come from claude --help, so no guessing. On the left is Codex made it separately—the function is correct but visually unforgettable. The right side shows the result of the same Claude Code session being reconnected and done with UI polish, and the bottom right corner says session: reused ✓ means no new Claude session was started. This is not a "all success" table, but a real chain of failures. The promise of v2.0.1 is not that Fable never fails, but that failure does not run endlessly, does not silently change modes, and does not treat half-finished results as a plan. The dollar amount here refers to the cost returned by the Claude CLI in the actual planning run, not the token savings rate for the entire workflow; If a failed attempt does not produce a final result, it honestly writes unknown. The actual task, model, effort, and successive cost for each identity can be found in the v2.0.0.0 failure baseline receipt and v2.0.1 complete conversation consumption receipt, respectively. Run boundaries are found in docs/releases/v2.0.1.md, references/bounded-planning.md, and docs/showcase-cost-model.md. When Codex reads this SKILL.md, it identifies itself as the host, runs references/codex-driven.md: orchestrates, implements, checks, and trims details; Claude Code is only responsible for planning, UI polish, and final /codex:review. The same Claude session is reused without repeated cold starts. Claude Code reads this SKILL.md and identifies itself as the claude_code, proceeding with references/claude-driven.md: planning, splitting, first passing the idea king's adversarial review, then handing over mechanical/quota-stressed tasks to delegate-codex.sh backend operations, using loops to track progress, and finally accepting a full review.

Overview (machine translated)

This page was translated automatically. Technical details may differ from the original.

Recommended Tools

Try a different keyword or remove a filter.

Install