parallel-minds
A Claude Code plugin with skills that leverage parallel agent swarms for creative exploration, ideation, problem-solving, and implementation verification.
Install
Via skills CLI
npx skills@latest add ogabrielluiz/parallel-minds
Pick the skills you want when prompted. Skills are installed globally and work with any coding agent that supports SKILL.md.
Via Claude Code plugin marketplace
claude plugin marketplace add https://github.com/ogabrielluiz/parallel-minds
claude plugin install parallel-minds
Skills
creative-consensus
Dispatches 10-20 parallel agents with domain-adapted creative angles to explore a design space, then presents structured proposals for you to choose from.
Invoke with /parallel-minds:creative-consensus or describe a creative challenge — the skill triggers automatically.
Recipes
| Recipe | Agents | When to Use |
|---|
standard | 10 | Default — feature design, refactoring |
deep | 15 | Architecture decisions, complex features |
thorough | 20 | System design, irreversible decisions |
research | 15 | When existing codebase context matters |
adversarial | 10 | When stress-testing matters more than breadth |
Domain Detection
Automatically detects the problem domain (systems architecture, API design, data modeling, security, DevOps, UX) and selects appropriate creative angles. Three mandatory roles always included:
- Regret Agent — "What will we regret about this approach?"
- Wildcard — Cross-domain angle nobody would think of
- Saboteur — Attacks the other proposals
Output
Comparison table in three tiers:
- Conservative — proven, low risk
- Moderate — novel but grounded
- Ambitious — high risk, high reward
Each idea includes: name, pitch, mechanism, effort, risk, reversibility, and failure mode.
implementation-scrutiny
Dispatches parallel verification agents to investigate whether an implementation is correct. Each agent takes a different verification angle — reference implementations, empirical tests, invariant checks, documentation comparisons.
Invoke with /parallel-minds:implementation-scrutiny or triggered automatically when code produces wrong results and initial fix attempts fail.
Modes
| Mode | Agents | When to Use |
|---|
fast | 3 core | Default — focused problem, clear domain |
full | 6-8 (core + domain) | Complex, multi-layered, or core agents disagree |
auto-escalate | 3 → 6-8 | Start fast, escalate if agents conflict or leave gaps |
Core Agents (always dispatched)
- Reference Hunter — Finds known-correct implementations and documents behavioral differences
- Empirical Tester — Writes and executes standalone tests with known inputs/outputs
- Invariant Auditor — Verifies each extracted invariant independently from first principles
Domain-Specific Agents
Activated based on detected domain: numerical/dsp, web-backend/distributed, frontend, data/ml, database, security, or general.
Evidence Standard
Every finding must include a validatable artifact — an executable script or a verifiable URL. The calling agent validates every artifact before accepting it. Findings without valid artifacts are demoted to UNVERIFIED.
Output
Structured findings report with severity tiers:
- Proven Bugs — script reproduced the issue, with repro recipe
- Likely Bugs — reference disagrees but untested
- Agent Disagreements — conflicting findings shown side-by-side
- Coverage Map — what was checked and what wasn't
Validated test scripts persist to scrutiny/ for regression coverage.
inbox-loops
Experimental. Relies on undocumented Claude Code behavior (claude --bg background sessions and CronCreate scheduling). Crons are currently session-only, so loops do not survive a restart without re-running setup, and interfaces may shift between Claude Code versions.
Bootstraps a multi-session "inbox" task system. Producer loops (PR, Jira, Vuln) each scan a source on a staggered cron and emit "needs action" tasks into shared queue files. A dispatcher loop routes each task to an auto-dispatched consumer subagent or escalates it to the user via push notification. Every loop runs as its own cron-armed background session and self-re-arms near the 7-day cron expiry to stay live indefinitely.
Invoke with /parallel-minds:inbox-loops, or describe intent to set up, start, check status of, or stop the inbox loops.
| Loop | Emits into | Default cadence |
|---|
| PR | inbox-prs.md | 13 */3 * * * |
| Jira | inbox-tickets.md | 43 */3 * * * |
| Vuln | inbox-vulns.md (sensitive, local-only) | 33 */6 * * * |
| Dispatcher | routes the three inboxes | 23 * * * * |
