🎸 RockstarRouter

English · Français · Español · Deutsch · Português · 简体中文 · 日本語 · 한국어

Stops your agent from booking a stadium tour to change a button color.

You installed a skill pack. Great songs, great band. Then you ask for a typo fix and the drummer counts in a 45-minute prog rock opera: brainstorming, spec, plan, tasks, review. Your tokens leave the building before the first chord.

RockstarRouter is the stage manager. It hears the request first and decides the setlist: a two-minute punk track, or the full stadium show.

🎚️ The setlist

| Lane | When | What happens | |---|---|---| | 🎸 QUICK | small, clear, reversible | locate, change, verify. No skill, no ceremony. | | 🥁 STANDARD | a real bug, or a few files | one focused skill at most (debugging, TDD) | | 🎤 DEEP | schema, migration, auth, payment, cross-module, new capability | your pack's full workflow, pyrotechnics included | | 🎧 CLARIFY | nobody knows what "done" looks like | two questions max, then back on stage |

The heavy artillery is never confiscated. The moment a request deserves it, RockstarRouter hands the mic to your pack's own workflow. And if a QUICK run starts growing (more files, a risky zone, a test failing somewhere else), it stops the song and re-routes.

When it does kill an overkill, it tells you:

🎸 superpowers:brainstorming wanted a full set-list meeting for one button. I just plugged in and played.
🎸 BMAD had a whole build pipeline warmed up for one button. Left it in the green room.
🎸 ecc:orch-add-feature was ready to run a full research-plan-TDD-review set for one button. I sent it home.

The attitude stays in the chat. Code, commits and PRs stay sober. Prefer a quiet roadie? ROCKSTAR_ROUTER_TONE=sober.

🤘 Plays with any band

No pack-specific code. RockstarRouter reads the skills you have installed and classifies them from their descriptions: anything that brainstorms, specs, plans or chains other skills only plays DEEP; debugging and TDD play STANDARD; anything it can't read is treated as DEEP and never gets a QUICK slot.

Packs that force their workflow on every request ("you MUST invoke a skill if there is even a 1% chance") get spotted and overruled for QUICK requests.

Tested on stage with:

| Pack | Fires heavy workflows on its own? | Profile in RockstarRouter | |---|---|---| | superpowers | yes, often | yes | | BMAD Method | sometimes | yes | | Spec Kit | no | no, generic rules | | ECC | no | no, generic rules | | OpenSpec | no | no, generic rules | | cc-sdd | no | no, generic rules | | Compound Engineering | no | no, generic rules | | Agent OS | no | no, generic rules |

Six of the eight packs ran on generic rules only, which is the real test.

📊 The numbers

14 requests on a small demo repo, 8 packs, each run with and without RockstarRouter. Claude Code 2.1.308, Sonnet, headless. Raw data in bench/results-2026-10-22.json.

Heavy workflow fired when a lighter path was enough (color, typo, log level, date format, small button, duplicate search results, CSV export, slow tests, two vague requests):

| Pack | Without | With RockstarRouter | |---|---|---| | superpowers (2 runs) | 3 and 5 times, once for a button color | 0 | | BMAD Method | 1, a build workflow on a CSV export, stuck, nothing shipped | 0 | | the 6 other packs | 0 | 0 |

Heavy workflow when it is warranted (rename a DB column, multi-currency, sessions to JWT, explicit /deep):

| Pack | Without | With RockstarRouter | |---|---|---| | superpowers | 2/4 | 4/4 | | BMAD Method | 2/4 | 4/4 | | Spec Kit, ECC, OpenSpec, cc-sdd, Compound Engineering, Agent OS | 0/4 each: multi-currency coded straight into 5 to 6 files, no spec | 4/4 each |

"Warranted" counts the pack's full workflow, or a stop to ask before a large change. The column rename on OpenSpec and cc-sdd needed a rule fix in v0.6 ("schema is always DEEP"), re-tested twice on 4 packs: 8/8.

The second table is the surprise of the tour. Most packs never pick up their own heavy instruments when the song calls for them. RockstarRouter does it for them.

What it costs: on trivial requests, from +3k to +57k tokens per request depending on the pack, on top of a 110k to 280k baseline. When RockstarRouter sends a big request to the full workflow, that workflow can cost a lot (cc-sdd spent 1.2M tokens specifying multi-currency). That is the stadium show you asked for.

Fine print: one model, one small repo, one or two runs per config. In headless mode a brainstorming skill stops at its first question, so the "without" runs look cheaper than a real interactive session. With no pack installed, the encore line still slipped out once in six runs.

Play it again yourself (needs git, node, npx, uvx and Claude Code):

cd bench && ./setup.sh
for id in $(seq 1 14); do for cfg in sp sprr bmad bmadrr spk spkrr ecc eccrr osp osprr ccsdd ccsddrr ce cerr aos aosrr; do ./run.sh $cfg $id; done; done
node parse.js runs && node score.js results.json

🎫 Backstage pass

Step 1: get the skill (pick one)

| Agent | Command | What you get | |---|---|---| | Claude Code (recommended) | /plugin marketplace add GregoryBabonaux/RockstarRouter, then /plugin install rockstar-router@rockstar-router | Skill and hooks: routing rule on stage at session start, /deep and /quick | | Any agent (Codex, Cursor, Gemini CLI, Copilot, OpenCode…) | npx skills add GregoryBabonaux/RockstarRouter | Skill, in the project or globally with -g | | No skills.sh | npx rockstar-router-cli skills --yes | Skill copied to ~/.agents/skills and ~/.claude/skills |

Step 2: put the routing rule in your project

Agents without hooks need the rule in their instruction file, otherwise the router only plays when the agent feels like loading it, and a pack that forces its own workflow usually grabs the mic first.

• npx rockstar-router-cli init shows what it will write. Nothing is written until you confirm.
• npx rockstar-router-cli init --yes writes the rule into AGENTS.md (and CLAUDE.md or GEMINI.md if your project uses them), between markers, without touching the rest of the file.

With the Claude Code plugin, this step is optional: the hooks already bring the rule on stage.

Step 3: soundcheck

• npx rockstar-router-cli doctor

Updates

| Installed with | Update | |---|---| | Claude Code plugin | claude plugin update rockstar-router@rockstar-router, or turn on auto-update in /plugin | | skills.sh | npx skills update (npx skills update -g for global) | | npm | npx rockstar-router-cli@latest update |

🎛️ Controls

Just ask. To override:

• /deep or deep: at the start of your message: full workflow, whatever the size;
• /quick or quick:: no workflow, straight to the riff.

Without the Claude Code plugin, deep: and quick: rely on the agent noticing them.

| Variable | Values | Default | |---|---|---| | ROCKSTAR_ROUTER_TONE | rock, sober | rock |

🩹 Known limits

• Routing is a model decision guided by a prompt, not a hard gate. It will sometimes misjudge.
• Some packs force their workflow through mechanisms a skill can't see. The router names what it skips so you can spot conflicts.
• Benchmarked on Claude Code with Sonnet only. On Codex CLI, we checked that the skill and the AGENTS.md rule load into the model context (codex debug prompt-input), but the routing behavior itself is not benchmarked there yet. Results on other agents and models welcome.

License

MIT. Turn it up.