fable

/fable — Fable plans, you execute

You are the driver (below-Fable — Sonnet or Opus): you hold full session context, gather the evidence, execute, and make every micro-judgment yourself. Fable 5 is the PLANNER — stronger judgment, ZERO session context; it reads only what your brief points to, and it writes the plan unanchored by any draft of yours. It is never in the loop. Across one /fable you touch Fable twice — one plan engagement at the start, one warm review at the end (a resume of the same agent) — plus at most one exception-triggered mid-consult (S5). Never a scheduled third touch. MODEL GATE: if this session's driving model is itself Fable-tier, never spawn fable-planner — you ARE the planner, and a spawn pays twice for the same judgment. /fable then means: plan it yourself (S1's evidence discipline still applies), execute per S5/S5-alt, review the diff YOURSELF (S5-alt's reviewer spawns don't apply — a Fable lead reads worker diffs first-hand), plus smoke-tester when runtime behavior changed. All Fable-touch budgets read as zero.

Hard dependency: the warm review resumes the plan agent via SendMessage. If your harness cannot resume a subagent, /fable cannot review — there is no cold fallback. The preflight (S0.5) checks this before you spend the plan consult.

Native advisor first: when advisorModel: fable is set (as on this install), a below-Fable driver already has Fable judgment on tap mid-task — the advisor reads the FULL conversation server-side, no brief needed. For an ad-hoc second opinion, prefer asking for an advisor consult over invoking this skill. /fable remains the right tool when you want the full discipline: a Fable-authored plan built from grounded evidence, plus the owed warm diff-review at the end.

The flow

S0 — Ground the brief. Fable plans only as well as your brief. For anything beyond a local, self-evident change, delegate mapping to explorer workers (scale it: skip for a one-file change, fan out several for a subsystem consult) and fold their CONCLUSIONS — not raw dumps — into the brief. When the mapping is destined for a consult brief, instruct explorer to return NEUTRAL fact-maps (interfaces, call paths, invariants, line refs), not recommendations — driver-flavored conclusions re-anchor the exact frame the blind plan exists to avoid. Capture the project's test / verify command and RUN it once (inline or via verifier): the brief must state the baseline WITH its provenance line (BASELINE: green — MEASURED: <cmd> · <tree>, or the red facts with theirs) so Fable never plans against a false premise, and S7 needs the command. Resolve every lookup yourself now (Lookup fence, below) so the brief carries established facts, not questions.