superpowers-reasonix

Process skills that give a Reasonix coding agent the disciplines of a careful engineer: write the test first, debug from evidence, plan before coding, verify before claiming done. Each one loads at the moment it applies.

A port of obra/superpowers for Reasonix (the DeepSeek-Reasonix CLI) and DeepSeek models. Same disciplines, rewritten for Reasonix’s tools, subagent model, and skill format. It is not a 1:1 transliteration. Adaptations covers what changed and why.

The skills

Eleven skills under skills/, covering a project from idea to merge:

Subagent dispatch, code review, parallel work, and codebase exploration aren’t skills here. Reasonix ships those as native tools (task, review, wait, explore), and the skills use them directly. Why?

Install

One-liner (recommended). Downloads the latest release and installs the skills into your global Reasonix skills root, where they auto-load in every session:

curl -fsSL https://raw.githubusercontent.com/christopherarter/superpowers-reasonix/main/install.sh | bash

Re-run anytime to upgrade — it skips if you’re already on the latest (a skip creates no backup; --force to reinstall), and any earlier manual install it finds is backed up before being replaced (nothing is silently destroyed). Pin a release with --version v1.0.0. Skills land in ~/.reasonix/skills/ (override with REASONIX_SKILLS_DIR). For the always-on discipline, copy the bundled AGENTS.md into each project:

cp ~/.reasonix/skills/.superpowers-reasonix/AGENTS.md /AGENTS.md

How a session uses them

brainstorming → writing-plans → executing-plans (inline)
                      └ or, for higher quality, dispatch per task with
                        the native `task` tool:
                            per task:  task   → implement + test
                                       (verify it matches the spec)
                                       review → code-review the diff
                            finishing-a-development-branch

test-driven-development, systematic-debugging, and verification-before-completion run throughout; using-git-worktrees sets up isolation up front.

Always-on discipline

Claude Code force-injects a using-superpowers skill at session start via a hook. Reasonix hooks can’t: SessionStart stdout never reaches the context, so a skill can’t establish discipline meant to apply before anything is invoked. Two always-on layers do the job instead.

1. The pinned skills index (automatic). Every skill with a description appears as one line in the system prompt, so the model finds them with no hook.
2. AGENTS.md (loaded every session). It holds the “load a skill before you act” rule, the situation-to-skill routing table, the rationalization red-flags, and the instruction-priority hierarchy. Benchmarking on deepseek-flash found this lifts skill invocation versus relying on a skill the model must choose to load.

Benchmarks

Two benchmarks under bench/, both targeting deepseek-flash, DeepSeek’s smallest coding model. A skill that fires reliably on the weakest model does at least as well on the larger deepseek-pro (the v4 Pro tier), so testing against the floor lets this port claim it works across the range. Pass --model=deepseek-pro to check a skill on the larger model.

Invocation measures which skill fires. bench/bench.mjs runs a structural gate over every SKILL.md, then sends realistic prompts through reasonix run and reads back the chosen skill. The 13-case corpus (eleven positives, one per skill, plus two negatives) passes 13/13 on deepseek-flash: every positive invokes its expected skill and neither negative fires. The expected skill comes first in 10 of the 11 positives; the branch-finishing prompt invokes superpowers-verification-before-completion before superpowers-finishing-a-development-branch, since “tests pass, let’s wrap up” legitimately cues verification first.

Execution fidelity measures whether the model obeys a skill already in context. bench/exec/ force-loads each skill in an isolated workspace and scores the transcript against a rubric: hard rules (test-before-implementation, ran-verification, no-placeholder-output) mechanically, softer criteria by a Claude judge. Save a baseline, edit a skill body, then re-run; any score that drops flags a body that lost discipline.

Run instructions and how to add cases are in bench/README.md and bench/exec/README.md.

Design notes

These skills supplement Reasonix’s native tools rather than compete with them. Reasonix already ships subagent orchestration: task (dispatch a subagent), review (code-review a diff), wait (join parallel jobs), and explore (investigate the codebase). So this port ships no orchestration or review skills. An earlier draft had subagent-driven-development, requesting-code-review, dispatching-parallel-agents, and worker subagents; benchmarking showed the model preferred the native tools, so they were retired. What’s left is the discipline the runtime doesn’t provide.

Only read-only work parallelizes cleanly, so fan out explore/read tasks and integrate sequentially. Rather than ship a separate subagent-driven skill, writing-plans folds the per-task discipline (implement, check the spec, then review) into instructions for the native task tool.

Staying current

This is a port, so it drifts from two upstreams: obra/superpowers (content) and DeepSeek-Reasonix main-v2 (the harness). UPSTREAM.json pins the commit last synced from each.

Run /superpowers-sync-upstream to check both, re-port what changed through the Adaptations rules, and re-pin once bench/ passes. A weekly GitHub Action (.github/workflows/upstream-drift.yml) does the detection on its own: when either upstream advances past its pin, it opens one upstream-drift issue listing the changed skills or contract files.

Adaptations from Claude Code superpowers

Reasonix resembles Claude Code but diverges in ways that break a naive copy. The disciplines themselves are preserved (the Iron Laws, red-flag tables, rationalization counters, RED-GREEN-REFACTOR); only the mechanics are rewritten.

License & attribution

MIT. See LICENSE.

Original concept and content: Jesse Vincent (obra/superpowers), MIT, Copyright (c) 2025 Jesse Vincent, preserved in LICENSE. This is a community port for the Reasonix platform (Copyright (c) 2026 Chris Arter).

Versioning: released as semver git tags (vX.Y.Z) on Releases; the installer pulls the latest. The skill set is the versioned surface — retiring/renaming a skill is a minor or major bump, adding one a minor, body/wording refinements a patch. bench/BASELINE.json is meaningful relative to the commit that captured it.