Superpowers Now Installs Into 16 Coding Agents, and Its Issue Tracker Shows the Cost

Superpowers, Jesse Vincent's MIT-licensed skills framework for coding agents, now lists install routes for 16 harnesses, from Claude Code and Codex CLI to Cursor, Gemini CLI and Qwen Code, according to its GitHub README. The repository page showed about 308,600 stars when fetched on Oct. 30, a figure The Terminal saw in one place only.

The Terminal did not install or run it. This piece draws on the README, the project's issue tracker and Simon Willison's posts on Vincent's work.

What the framework makes an agent do

The README calls it "a complete software development methodology for your coding agents," built by Vincent with the team at Prime Radiant. It starts by clarifying intent, then writes a plan, then runs implementation through subagents with reviews. The README's wording on enforcement is "Mandatory workflows, not suggestions" and "The agent checks for relevant skills before any task."

The core sequence has seven skills, in this order: brainstorming, using-git-worktrees, writing-plans, subagent-driven-development (or executing-plans), test-driven-development, requesting-code-review and finishing-a-development-branch. Test-driven development means the agent writes a failing test first. The README says "Critical issues block progress" at the review step.

Vincent announced it on Oct. 9, 2025, a few days after Claude Code added plugins. Willison summarised it on Oct. 10, 2025 as Vincent "wrapping up a whole host of his accumulated tricks as a new plugin." Vincent told readers then that "The core of it is VERY token light," pulling in one document of fewer than 2,000 tokens.

Where the issue tracker disagrees

The tracker showed 194 open issues and 178 open pull requests on Oct. 30. Several concern the two things the framework promises, which are low cost and fixed order.

Issue	Opened	Report
#2612	Oct. 23, 2026	Porting a roughly 2,000-line R script with subagent-driven development hit five session limits; the coordinator's context reached 830,000 tokens in Claude Code
#2638	Oct. 28, 2026	In Antigravity, a multi-step cleanup ran after a diagnostic recommendation without invoking writing-plans
#2628	Oct. 25, 2026	In Cursor, a fix request in the same turn as the skill was left after an intake question

The first is a long-session cost, and the quoted figure comes from the reporter in the issue, not from a measurement The Terminal could repeat. It does not contradict the "fewer than 2,000 tokens" claim, which was about the startup document, but it shows that the subagent loop is where the budget goes. The Terminal reported a related overhead question in Claude Code's background agent status line.

The second and third matter because the README says the workflows are mandatory. Both reports come from non-Claude harnesses, which fits the project's own rule that skill updates must work across all supported agents. Issue #2567 describes one cause: brainstorming's builder check has no fallback for a harness that only lets an agent spawn subagents when the user asks.

The telemetry the README discloses

The README says the brainstorming visual companion loads the Prime Radiant logo from the project website, and that the request includes the Superpowers version. It says the request leaves out project, prompt and agent details. Setting SUPERPOWERS_DISABLE_TELEMETRY to a true value turns it off, and the project also honors Claude Code's DISABLE_TELEMETRY and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC settings.

A diagnosing-superpowers skill reads session transcripts and can export a scrubbed bundle for bug reports. The README does not say how the scrubbing works. No open issue in the list The Terminal read mentions telemetry.

Who is allowed to change it

The contribution rules are narrow. The project says it generally does not accept new skills, that skill updates must work across all 16 harnesses, and that agent-written pull requests for issues filed by others are not accepted unless a maintainer asks. With 178 pull requests open, that policy is a queue.

An agent-heavy review workflow carries its own review cost, which The Terminal examined in Brooks's law and AI agents. Superpowers adds a requesting-code-review step by design, so the tracker's skipped-step reports and its token report pull in opposite directions: fewer steps are cheaper and more steps are safer.

The README shows no release number or date in the captured page, so a team pinning a version has to pin a commit. The number to watch is issue #2612, since it is the one a team would hit before deciding the framework fits a large port.