octopus-skill

Umbrella entry for the octopus-skill prompt library — pick the right arm for a long-horizon agent task and follow it. Two arms share one discipline (verified-not-written "done", no test theater, no speculative building, forced convergence, owner red lines). The `loop-graph` arm builds an executor + clean-context supervisor as two scheduled loops — for the `/loop` primitive (Claude Code, grok, Cursor, shell), multi-milestone phases, executor/supervisor split across hosts/models, owner-gated runs. The `quest` arm emits one objective prompt handed to a host that already drives a goal to done with its own verifier — for the goal primitive (grok `/goal`, a Codex task). Use this when the user invokes /octopus or is unsure which arm fits.

Run a long-horizon coding task as a small graph of agent nodes driven by two scheduled loops, instead of one drifting loop. Generates an executor node (works against a single-source-of-truth ledger) and a clean-context supervisor node that audits from outside the executor's context — re-verifying claimed-done work, checkpoint-committing what passes, and correcting drift through a one-way directives file. Use for multi-round tasks where an agent tends to scope-creep, fake "done", or quietly lower the bar, and where you'll drive it with the `/loop` primitive (Claude Code, grok, Cursor, shell). Nodes share no context, so the executor runs on a cheap/fast model while a strong model supervises. (The methodology this arm implements is still called "graphkit" internally; runs generate into `.graphkit/<date-slug>/`.)

Compile a long-horizon coding task into ONE objective prompt for a host that drives a goal to done with its own verifier (grok `/goal`, a Codex task) — where the host's harness is the acceptance auditor, so no second supervisor loop is needed. Folds the octopus discipline (verified-not-written "done", no test theater, no speculative building, forced convergence, owner red lines) into the objective text. Use for a single self-contained goal on a goal-capable host; for multi-milestone / cross-host / owner-gated work, or a host that only loops (Claude Code, Cursor, shell), use the loop-graph arm instead.