fable-harness

Start substantial coding work on Claude Fable 5.1. On a project with no spec, it creates SPEC.md from the request and whatever structures the user supplies, then stops for review. On a project with a spec, it briefs the next slice (outcome, checkable done-conditions, out-of-scope, assumptions, checks, branch), runs it to completion, verifies, and updates the spec. Use at the start of any task beyond a few tool calls (features, multi-file changes, refactors, migrations, debugging), or when the user says "brief this", "start the project", "next slice", or "run this end to end".

Save transient session state when stopping mid-slice so the next Claude Code session resumes exactly where this one stopped. The spec holds the durable expectations; the checkpoint holds what the spec does not — position within the current step, problems and how they were handled, options tried and set aside, and details that are hard to reconstruct. Use before stopping mid-task, when the user says "checkpoint", "save state", "hand off", or "we'll continue later".

Set or report the harness's provider mode for this project. `fable` (default) has Fable 5.1 implement each slice itself; `codex` hands implementation to an OpenAI Codex model (gpt-5.6-sol or gpt-5.6-luna) through the user's installed `codex exec`, routed by task class, while Fable briefs, keeps the spec, and verifies. Use when the user says "switch to codex mode", "use sol/luna for implementation", "what mode are we in", or "back to fable mode".

Independently verify a finished slice in a fresh, read-only context before reporting it done — checks the diff against the spec's done-conditions, runs the project's checks, hunts for bugs, and records verified status in the spec. Use after completing a slice, before opening a PR, or when the user asks "is this actually done", "verify", or "double-check".