Spectra

Spectra

[](LICENSE) [](https://github.com/dnhess/spectra/actions/workflows/lint.yml)

A local-first multi-agent deliberation runtime for structured review and debate using a **blackboard architecture** — every perspective refracted, every angle examined. Claude Code remains the default execution path; the provider-neutral runtime now also supports an opt-in, approval-gated Codex Quick peer-review opening phase.

Available Skills

deep-design

**Rigorous multi-perspective design review.** Use when a document, spec, or idea needs stress-testing from every angle before implementation. Produces a revised document with all findings incorporated.

Triggers: design docs, architecture specs, product requirements, feature proposals.

decision-board

**Structured multi-perspective debate.** Use when a decision needs structured debate before committing. Produces an Architecture Decision Record (ADR) with recommendation, dissent, and conditions.

Triggers: architectural decisions, technology selection, build-vs-buy, migration strategy.

peer-review

**Multi-perspective code review.** Use when code changes need thorough review from multiple specialist viewpoints before merge. Includes a Scout + Research reconnaissance phase that gathers codebase context and best practices before reviewers begin.

Triggers: pull requests, code changes, refactoring review.

trust-layer

**Adversarial verification for AI-generated output.** Use before accepting any AI-generated code, diff, file, or Spectra session recommendation. Four adversarial personas (Package Validator, Intent Auditor, Security Challenger, Coherence Checker) challenge the output from independent angles.

Triggers: AI-generated code, diffs, files, or Spectra session output needing verification before acceptance.

coherence-monitor

**Metacognitive audit for long-