Reverse Engineer Orchestra

An open, portable cognitive-engineering agent for Codex and Claude Code.

It works backward from an observable result, reconstructs the system behind the request, compares competing explanations, chooses the smallest specialist team, builds the artifact, and verifies it before claiming success.

This repository is not model weights and does not manufacture a human biography or consciousness. It is a practical agent system: cognitive functions, bounded roles, executable control gates, reviewed knowledge modules, cross-platform skills, and release tests that make a capable language model work in a more disciplined and reproducible way.

Русская инструкция

Requirements

• Python 3.11 or newer. The specialist factory uses Python's built-in tomllib, which is unavailable in Python 3.10 and earlier.
• Codex and/or Claude Code.
• No third-party Python packages are required.

Check Python before installation:

python3 --version

On Windows, use py --version. If it reports a version below 3.11, update Python before running the installer.

What is included

• 15 cognitive functions and an INTAKE -> ... -> VERIFY -> STOP loop;
• a 58-capability orchestra instantiated only when the task needs it;
• a deterministic Python control engine with review and memory gates;
• a specialist factory that creates bounded, unproven-by-default skills;
• brand, ethical offer, audience, and experimentation systems;
• original cultural-editorial writing informed by a rights-conscious public-metadata-and-feature study;
• a social-psychology hypothesis engine with 20 cross-cutting, source-bounded mechanism cards;
• 14 non-sequential practitioner cards informed by an audit across all 32 public course chapter ranges;
• creative landing-page direction with accessibility, performance, responsive, and reduced-motion requirements;
• source research, fact checking, and GitHub archaeology;
• native Codex plugin and Claude Code adapters;
• an offline installer and a release doctor with no production dependencies.

The eight Skills

Skill	Responsibility
reverse-engineer-orchestra	Reconstructs the outcome, preserves competing explanations, selects specialists, executes, and verifies.
brand-systems	Builds evidence-led positioning, audience, messaging, ethical offer/funnel, and brand experiments.
editorial-intelligence	Creates original culture-aware articles, scripts, headlines, and landing copy with factual and originality gates.
social-psychology-engine	Turns claims about social behavior into bounded mechanisms, countermodels, predictions, and tests.
creative-landing-pages	Directs and builds authored responsive landing pages, including justified motion/3D and accessibility/performance review.
source-research	Produces current, traceable evidence packets and provenance records.
fact-checking	Audits claims, citations, recency, contradictions, and evidence labels.
github-archaeology	Evaluates repositories, dependencies, licenses, maintenance, security, and architecture candidates.

The first Skill is the conductor. It loads or delegates to the other seven only when the task needs them.

Fastest start from a downloaded ZIP

1. Download and extract the repository.
2. Open a terminal in the extracted folder.
3. Confirm Python 3.11+ and check the downloaded package:

python3 --version
python3 scripts/orchestra.py doctor

4. Install for both Codex and Claude Code:

python3 scripts/orchestra.py install both --scope user

5. Run the doctor from each installed location:

python3 "$HOME/.agents/scripts/orchestra.py" doctor
python3 "$HOME/.claude/scripts/orchestra.py" doctor

6. Restart the application. Invoke it with:

• Codex: $reverse-engineer-orchestra
• Claude Code: /reverse-engineer-orchestra

On Windows PowerShell, the equivalent commands are:

py --version
py scripts\orchestra.py doctor
py scripts\orchestra.py install both --scope user
py "$HOME\.agents\scripts\orchestra.py" doctor
py "$HOME\.claude\scripts\orchestra.py" doctor

If only one platform is installed, run only its installed doctor. READY means the installed core Skills, cognitive engine, and specialist registry passed their local checks. It does not certify the quality of future outputs.

Codex plugin installation from GitHub

codex plugin marketplace add kirilenkofedir-max/reverse-engineer-orchestra
codex plugin add reverse-engineer-orchestra@reverse-engineer-orchestra

Restart Codex, then start a task with $reverse-engineer-orchestra followed by the real outcome you need. The ZIP installer above is the recommended route when you want a directly inspectable local install and post-install doctor; Codex manages a marketplace plugin's storage location internally.

Claude Code installation

Either open the cloned repository in Claude Code—the checked-in .claude/ adapter is discovered as project configuration—or install it globally:

python3 scripts/orchestra.py install claude --scope user
python3 "$HOME/.claude/scripts/orchestra.py" doctor

Restart Claude Code and invoke /reverse-engineer-orchestra.

How to use it

Do not write a perfect brief. Give it the unresolved result and real limits:

$reverse-engineer-orchestra
BUILD: I need an original launch landing page for this offer. Reconstruct the
strategy, research what must be current, use the editorial and creative-web
systems, build it, and verify mobile, accessibility, performance, and claims.

$reverse-engineer-orchestra
DEBUG: People reach checkout but do not complete it. Diagnose before changing
the funnel. Separate observation from explanation and propose the cheapest test
that can distinguish the live hypotheses.

The available modes are ANALYZE, DESIGN, BUILD, DEBUG, and AUDIT. The agent chooses FAST, STANDARD, or DEEP depth based on ambiguity and risk; you can request one explicitly.

Version 1.1 uses visible, real subagent threads for material work. Public copy runs through a maker and a distinct reviewer. A request that combines niche or market research with creation runs research-scout and analyst first, then maker and reviewer. If custom profiles are unavailable, the route falls back to Codex's built-in explorer, worker, and separate default agents. If the host has no subagent tool, the conductor must disclose a single-agent fallback instead of claiming that an orchestra ran.

Auditable runtime

Normal work uses the instruction layer. For a persistent high-scrutiny run:

python3 scripts/orchestra.py describe
python3 scripts/orchestra.py engine start \
  --objective "Observable result" \
  --mode BUILD \
  --depth DEEP \
  --risk reversible \
  --acceptance "Named verification check"

The engine stores local run state under knowledge/reverse-founder/; generated databases and run archives are ignored by Git. Its canonical case memory ships empty. Only separately reviewed cases may be promoted.

Source and originality boundaries

The package contains derived mechanisms, audits, and original procedures—not the source book, a video transcript, copied magazine articles, page images, or a named writer's voice. It is not affiliated with DAZED, David G. Myers, McGraw-Hill, Posts & Telecom Press, Caleb Ralston, YouTube, Anthropic, or OpenAI. See NOTICE.md.

Verification

python3 scripts/sync_package.py
python3 scripts/validate_package.py

The release gate checks skills, JSON/JSONL, adapter parity, private markers, forbidden source files, and the cognitive-engine test suite. See docs/ARCHITECTURE.md and docs/VERIFICATION.md.

A live Codex maker/reviewer correction cycle is recorded in docs/LIVE_CODEX_BEHAVIOR_PROBE.md. Claude Code received clean-install, profile, Skill, and self-contained-module checks; a live Claude model run was not available on the release machine and is not claimed.

License

The original code and documentation in this repository are released under the MIT License. Third-party works and trademarks remain their owners'.