---

name: octopus description: 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, Codex, 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.

octopus 🐙 — one brain, many arms

A curated prompt library for long-horizon agent work. One shared discipline, compiled to whichever host you run. Your job on invocation: pick the arm, then read and follow that arm's SKILL.md. Don't reimplement it here.

Pick the arm — task shape decides; the host only gates what's available

Decide by task shape first, then check the host can run it (full matrix in lib/host-dialects.md):

Choose	When
quest	a single self-contained goal · reproducible acceptance · no non-skippable milestone gate · no owner-only call on the critical path → emit one objective prompt, run it with the host's goal command, ride its own verifier — no second loop
loop-graph	any of: multi-milestone phases with a non-skippable gate · executor and supervisor on different hosts/models/cadences · owner red-lines/gates that must stop the run · cross-session durability · you want an independent verifier → build the executor + clean-context supervisor two-loop graph

The host only gates which arms are reachable — most now do both:

Host	Can run	How
Grok	both	quest rides its native adversarial verifier; loop-graph = two /loops
Codex	both	quest = /goal / send-as-task (self-drives to done). loop-graph = drive both nodes with an interval /loop (heartbeat), never /goal — a goal harness re-fires a parked/stalled node forever (livelock), having no "waiting for a directive" rest state. Codex /loop needs an explicit interval (e.g. 4m) or no timer is created.
Claude Code	both*	quest = a self-paced single /loop (*no /goal command, no independent verifier); loop-graph = two /loops (its supervisor is the verifier)
Cursor, shell/cron	loop-graph only	no goal primitive

Default when genuinely unsure: loop-graph (it carries its own verifier and runs on every loop-capable host).

Then

Read the chosen arm's SKILL.md and run its interview → generation → delivery. Shared reference both arms rest on:

• lib/host-dialects.md — per-host /loop · goal · hook syntax, which primitive each host has, and wake/notify primitives.
• lib/methodology.md — why each discipline rule exists and the failure mode it prevents.