megapowers

megapowers is exactly one plugin for current Claude Code and Codex. It adds
task-level skills (the inventory lives in plugins/megapowers/skills/catalog.json),
one shared communication style, and a small destructive-command
tripwire. It does not replace native agents, plans, goals,
permissions, worktrees, memory, or browser tools.

The plugin is deliberately small. There is no model router, session catalog,
formatter, status line, custom scheduler, hosted service, or compatibility
layer for other harnesses.

Install

Claude Code:

claude plugin marketplace add lawzava/megapowers@release
claude plugin install megapowers@megapowers

Codex:

codex plugin marketplace add lawzava/megapowers --ref release
codex plugin add megapowers@megapowers

Start a fresh session after installation. The release branch only moves to
signed release tags, so marketplace refreshes never install unreleased main
state. Full update, verification, pinning, and uninstall instructions are in
docs/install.md.

The workflow

Use the smallest skill that matches the task. Skill descriptions stay visible;
the full body loads only when selected.

Repository instructions, existing code, and configured project tools remain
authoritative. See docs/orchestration.md for the
native-first task shapes and execution map. It adds no scheduler or agent
runtime.

For non-trivial delegation, orchestrating can read one optional personal
capability registry from ~/.config/megapowers/agent-capabilities.md. The file
stays outside the plugin and supplies advisory choices only; it creates no
agent access or authority.

What installs

• Task skills and their maturity listed in skills/catalog.json.
• One shared style for direct, concise technical replies: a selectable Claude Code
  output style and a trusted Codex startup hook.
• One PreToolUse Go hook for obvious catastrophic commands.
• One Go standard-library independent-review tool, loaded only with that skill.

Select Megapowers through Claude Code's /config output-style picker. It
preserves built-in coding instructions and respects another selected style.
Codex adds the same style as developer context at session startup and after
compaction once the user trusts
the bundled hooks. Set MEGAPOWERS_OUTPUT_STYLE=off before launching Codex to
omit the style while keeping the guard and shared workflow guidance. Both
harnesses receive instructions to write new helper code in Go, including
temporary scripts and one-off commands, and use native tools for direct edits.
Neither adapter changes global user
configuration. Hooks require Go and cache their compiled executable locally.

Claude Code and Codex receive the same high-confidence denials. Reversible risk
remains with each harness's own permission system. The hook is an accident
tripwire, not a sandbox. Read
SECURITY.md before trusting the plugin.

Evidence

Four evidence classes are kept separate:

1. scripts/validate.sh and evals/run-all.sh run bounded deterministic
   regressions and runner selftests. They prove repository mechanics, not agent
   quality.
2. Trigger recall measures trace-proven skill selection. Its configured gates
   enforce Claude study results and report Codex results without enforcement.
3. The optional installed-plugin A/B study compares this checkout with an empty
   control under Claude Code and Codex through a hash-pinned isolation broker.
   It reports treatment reliability and paired control outcomes; it does not
   gate releases or claim that the plugin improves general model capability.
4. PR replay uses hidden correctness tests against pinned historical changes.
   Credentialed execution is disabled pending broker schema 2; the study is
   report-only.

No current-candidate behavioral result is claimed from selftests. Historical
measurements, including null results and their limitations, remain in
evals/RESULTS-archive.md. Current protocols and gates are in
evals/README.md and
docs/advanced/evals.md.

Limits

• Skills are instructions, not enforcement. Model and harness behavior can
  change.
• The communication style applies to the main conversation, not ordinary
  subagents, and it cannot suppress tool-result rendering. Codex skips the
  adapter until the user trusts the plugin hooks.
• The destructive guard matches a narrow set of command strings. OS sandboxing
  and least privilege remain the real boundary.
• Independent review discloses approved source content to the selected provider.
  It rejects common secret patterns, not every possible secret.
• Real-agent studies require credentials, spend, and a reviewed broker that
  keeps credentials and hidden state outside the actor's OS boundary. Their
  selftests do not substitute for a credentialed run.
• Only current Claude Code and Codex are supported. Exact structural and
  behavioral evidence boundaries are in
  docs/harness-support.md.

Develop

scripts/validate.sh
bash evals/run-all.sh --json results.jsonl

The deterministic suite fails on malformed, incomplete, indeterminate,
timed-out, or harness-error results. Contributions that change behavioral
guidance need deterministic regression coverage. Credentialed installed-plugin
A/B remains optional diagnostic evidence. See CONTRIBUTING.md.
The repository-local verification-map pilot is documented in
docs/advanced/verification-maps.md.

License and origin

megapowers is MIT-licensed. It began from Superpowers by Jesse Vincent and
retains other source credits in ATTRIBUTION.md. It is not an
Anthropic or OpenAI product.