Spec-driven development with learning feedback loops — installable into any repo in three commands.

AI-assisted development is fast but forgetful. The spec lives in a chat message, the reasoning behind a decision evaporates, and the same mistakes come back next week. Spectra fixes that by making intent explicit before you build and capturing learning after — all in version control, all driven by your coding agent.

It's packaged natively for Claude Code, OpenAI Codex, Gemini CLI, and Cursor — one source of truth, one set of commands — so any repo can adopt the entire protocol (the workflow, the review personas, the artifact structure, and a reflection reminder) without copy-pasting a thing.

Quick start

Install through your agent's native plugin system, then run /spectra-install in the repo you want to adopt it.

Claude Code

/plugin marketplace add rogueoak/spectra
/plugin install spectra@spectra
/reload-plugins
/spectra-install

/reload-plugins makes the newly installed plugin's commands (/spectra-install and the rest) available in your current session — without it the install won't have taken effect yet.

OpenAI Codex

codex plugin marketplace add rogueoak/spectra

Then install the spectra plugin from that marketplace and run /spectra-install.

Gemini CLI

gemini extensions install https://github.com/rogueoak/spectra

(or gemini extensions link . for local development), then run /spectra-install.

Cursor

Add the rogueoak/spectra marketplace (in-editor marketplace panel or /add-plugin), then run /spectra-install.

/spectra-install scaffolds your repo, drops in the protocol and review personas, installs a reflection hook, and points your AGENTS.md at it. Later, pull updates with /spectra-update.

The protocol

Every change flows through one loop (full text: spectra/protocol.md):

#	Step	What happens
1	Route	Trivial change? do it. New feature? write a spec. Bug or friction? write feedback (so it becomes a lesson).
2	Spec	Developer approves before any code is written (docs/specs/NNNN-*.md).
3	Plan	Multi-step work becomes an ordered plan (docs/plans/NNNN-*.md).
4	Build	Executed in a git worktree on a branch.
5	Test	Run the suite and fix until green before committing (no suite? add the test that proves the change).
6	Review	A PR is reviewed by the personas you've enabled (engineer, tester, architect, security by default — see Skills) who comment in a fixed format; major/blocker findings become learnings.
7	Merge	On approval.
8	Reflect	Before concluding, the living docs in docs/overview/ (project, features, architecture, learnings) are updated. A non-blocking pre-commit hook nudges you if you forget.

The feedback → learnings loop means the system gets better at your codebase over time, instead of repeating itself.

Skills

Spectra installs as a handful of slash commands (agent skills). Run them from the repo where Spectra is installed:

Command	What it does
/spectra-install	Adopt Spectra in the current repo — scaffolds docs/, copies the protocol and review personas, seeds the enabled-persona config, installs the reflection hook, and wires up AGENTS.md.
/spectra-update	Re-sync the Spectra-owned files to the installed plugin version (protocol, personas, host block, hook). Leaves your specs/plans/feedback/overview, your personas.config, and your user.md untouched.
/spectra-setup	Define your repo's 👤 User (ICP) review persona through a short guided dialog, so reviews can judge a change on your customer's behalf. Re-run to refine it.
/spectra-persona-enable [persona]	Turn on a review persona. With no argument, lists the personas available to enable as a numbered menu.
/spectra-persona-disable [persona]	Turn off a review persona (a core one too). With no argument, lists the personas currently enabled.

Review personas

Each PR is reviewed only by the personas you've enabled (tracked in docs/spectra/personas.config) whose facet the change actually touches — so reviews stay scoped, not eight bots on every diff. Four ship on by default, three more are available, and one is yours to define:

Persona	Default	Reviews for
🔧 engineer	on	correctness, edge cases, maintainability
🧪 tester	on	coverage, edge cases, honest tests
📐 architect	on	boundaries, dependencies, design-for-change
🔒 security	on	auth, input handling, secrets, dependencies
🎨 designer	off	visual consistency, spacing, design tokens, clear calls-to-action
⚖️ compliance	off	accessibility, PII minimization, i18n, GDPR/CCPA
📊 analytics	off	event tracking, measurable outcomes, feature-gate metrics
👤 user (ICP)	/spectra-setup	whether the change actually serves your ideal customer

Flip any of them with /spectra-persona-enable / /spectra-persona-disable. A disabled persona costs nothing — its checklist only loads when it's both enabled and scoped into a review.

Low token cost

Spectra is deliberately terse — the whole protocol fits in a corner of the context window, leaving room for your actual code:¹

What loads into context	Characters	Tokens (≈4 ch)
Always-on host block (in AGENTS.md)	693	173
Protocol only (no personas needed)	4,914	1,229
Full protocol + core personas	10,655	2,664
Optional personas (load only when enabled)	2,636	659
Everything, incl. install/update skills	26,677	6,669

What lands in your repo

docs/
  spectra/protocol.md      the protocol (agent reads this)
  spectra/personas/*       review lenses (all shipped; enable the ones you want)
  spectra/personas.config  which review personas are enabled (yours to edit)
  specs/                   approved specs (NNNN-<slug>.md)
  plans/                   ordered build plans (NNNN-<slug>.md)
  feedback/                bugs & friction → lessons (NNNN-<slug>.md)
  overview/                living docs, updated every change
AGENTS.md                  points your agent at Spectra
.git/hooks/pre-commit      reflection reminder (copied in; not tracked)

Repo layout (this repo)

• spectra/ — the shippable source of truth: the plugin (skills, protocol, personas, hook). This is what gets installed.
• docs/ — Spectra dogfooding itself: this repo built using its own protocol.
• assets/ — the logo and protocol diagram (SVG) used by this README.
• scripts/ — repo-local tooling, not shipped: token-report.sh keeps the token figures above honest (enforced by a pre-commit guard and test.sh).

License

See LICENSE.

Footnotes

1. Measured from the Markdown in spectra/ with a dependency-free ~4 chars/token heuristic, and kept in sync with the source automatically. ↩