Pause a heavy session at a NATURAL boundary and hand off to a fresh session via an audited handoff doc + ONE paste-able resume one-liner. Trigger when the user asks to resume/continue in a NEW session or chat ('resume in a new session', 'new session resumption', 'session handoff', 'let's pause and continue in a new chat', 'give me the resume one-liner', 'pause at a natural boundary and continue in a new session'), or asks to wrap up because the session/context is heavy. ALSO trigger proactively: when the session context has grown very heavy (roughly 500k+ tokens of history — multiple compactions, a long multi-phase transcript; a heuristic, not a counter) and a natural boundary is near, PROPOSE this skill to the user in one line; never pause on your own. Project-agnostic — any repo, any teammate, any session type. The handoff gets a MANDATORY clean-context audit (cross-model via Codex preferred when installed AND consented) to minimize context loss, and the deliverable is ONE line pasted into the new session.
Run token-disciplined Fable 5 sessions with the 10-80-10 model — use when starting any multi-step coding or research task, when work could fan out to parallel subagents, when the user mentions token cost, usage limits, or "save tokens", when the user asks to delegate or orchestrate work across models, or when the user asks to audit or clean up stale skills and CLAUDE.md left over from before Fable. Fable plans and reviews (the two 10s); cheaper Claude models (Opus, Sonnet, Haiku subagents) do the verbose middle 80%. Keeps Fable's expensive tokens on planning and quality ownership, not on work a cheaper model could do.
Use when deterministic/backend/algorithmic work needs an independent correctness check — numeric, boundary, parsing, state-machine, concurrency, or scoring/data math and API contracts; when critiquing a plan before multi-step deterministic execution; when reviewing a diff before committing; when you want a debugging second opinion after a failed fix attempt; or when the user mentions Codex, GPT-5.6, gpt-5.6-sol, GPT-5.5, or asks for a second model's opinion. Extends and requires fable-god-mode; needs a paid ChatGPT plan with the Codex CLI installed and logged in.