Skip to content

WookLabs/athanor

Repository files navigation

Athanor

License: MIT Claude Code

The alchemist's self-sustaining furnace — a workflow orchestrator that grows smarter with use.

v0.12.0 — 10 athanor-native commands + 1 KEEP vendored skill (/athanor:ce-test-browser) + 2 KEEP CE sub-agents. Clean-context workers. 3-tier adversarial planning. 6-lens parallel review. Sessions that compound. Thin Leader / cross-model adversarial planning / Spec-then-TDD discipline / Stop hook runtime gate are the athanor identity invariants — preserved intact through the v0.12.0 cutover. See CLAUDE.md §"Concept Absorption Surface" for the post-cutover surface and docs/v0.12.0-migration.md for the migration guide.

What v0.10.0 absorbed (corrected by v0.12.0 atomic cut)

  • compound-engineering 3.8.3 — originally vendored 33 skills + 49 sub-agents at v0.10.0. v0.12.0 atomic cut reduced this to 1 KEEP skill (/athanor:ce-test-browser, D8) + 2 KEEP sub-agents (ce-git-history-analyzer, ce-repo-research-analyst, D12). The remaining 32 skills + 47 sub-agents were removed in the v0.12.0 concept-kernel cutover. See docs/archive/v010-v011-vendoring-scope-correction.md for the plan-of-record misread retrospective.
  • superpowers 5.1.0 — originally vendored 13 skills at v0.10.0. v0.12.0 atomic cut reduced this to 0 surviving skills; 2 LIFT-source concepts (sp-systematic-debugging, sp-using-superpowers) absorbed into athanor-native skills as prose subsections, the rest dropped.
  • 5 concepts absorbed as native prose (NOT vendored directories): reviewer-persona vocabulary, doc-review mode, systematic-debugging Iron Law, requirements capture R/A/F/AE-IDs, skill-discovery preamble. Full MIT attribution in NOTICE.md §"Concepts adopted from upstream".

What v0.12.0 deliberately does NOT do

  • Does NOT silently downgrade /athanor:plan to a single-agent flow (cross-model adversarial dual-planner stays the default).
  • Does NOT silently downgrade /athanor:work to a non-TDD execution flow (athanor-native Spec-then-TDD stays the default).
  • Does NOT extend the Stop hook gate scope — per D10/D11 the runtime gate + companion-fix arc 5 layers + v0.10.2 normalization stack survive intact, with rationale re-framed to general defensive coverage.
  • Does NOT re-license athanor (athanor stays MIT; CE and superpowers stay MIT under their copyright holders).
  • Does NOT require CE or superpowers to be separately installed. Users who need a removed upstream skill install the upstream plugin directly — see the migration guide.

Why Athanor Exists

Without Athanor, your Claude Code session is a single brain trying to hold everything at once — the analysis, the plan, the implementation, and quality checks. Context fills up. Quality degrades. When you ask for a plan, the same model that writes code also evaluates it — no second opinion.

With Athanor, the main session never touches a file. It dispatches work to clean-context specialists — a researcher who brainstorms, a planner who architects, a critic who cross-reviews, an executor who implements with verification loops. By default, a single planner gets cross-model review from Codex. In deep mode, two independent planners compete and cross-review each other. Lessons from every session compound — future workers start smarter than the last.

Other tools give Claude more skills. Athanor gives Claude a team.

Quick Start

Prerequisites: Claude Code

In Claude Code:

/plugin marketplace add WookLabs/athanor
/plugin install athanor@athanor
/athanor:setup

Then try it:

/athanor:discuss "Add user authentication — OAuth vs session-based?"
For local development
git clone https://github.com/WookLabs/athanor.git
claude --plugin-dir /path/to/athanor

See It Work

You:     /athanor:discuss "Add a caching layer for the API"

         [dispatches researcher + devil's advocate in parallel]

         Researcher: 3 strategies — Redis, in-memory LRU, CDN edge cache
         Devil's Advocate: Redis adds ops complexity for a team of 1.
                          In-memory LRU handles current traffic fine.
         Critic: Start with in-memory LRU. Add Redis at 10K RPM.

You:     /athanor:plan

         [dispatches Planner (Claude)]
         Plan: Cache middleware, 4 phases
         [Codex reviews the plan]
         [Critic refines with review feedback]
         Final plan: 6 subtasks, dependency-ordered.

You:     /athanor:work --team

         Wave 1: [subtask 1, 2] parallel
           ↓ discovery relay
         Wave 2: [subtask 3, 4] depends on wave 1
           ↓
         Wave 3: [subtask 5, 6]
         All 6 subtasks complete. 2 lessons saved for next time.

Need deeper analysis? Try /athanor:deep-plan for full adversarial planning. In a hurry? /athanor:lite-plan skips the review step.

Commands

Command Mode What it does
/athanor:setup Health check and configuration
/athanor:discuss Plan Decision brainstorming (researcher + devil's advocate + critic)
/athanor:analyze Plan Parallel fast analysis (multiple workers simultaneously)
/athanor:debug Plan Triage-first parallel failure diagnosis
/athanor:deep-plan Plan Full adversarial planning (Claude + Codex cross-review)
/athanor:plan Plan Standard planning + review + task splitting (default)
/athanor:lite-plan Plan Lightweight planning (Claude only, no review)
/athanor:work Execute Grinding through every subtask until done
/athanor:review Plan Parallel 6-lens code review (architecture, quality, security, performance, testing, docs)
/athanor:discuss  →  /athanor:analyze  →  /athanor:debug (optional)
   "What?"              "Where?"              "Why broken?"

                  →  /athanor:plan (or deep-plan / lite-plan)  →  /athanor:work
                        "How?"                                       "Do it."

Everything before /athanor:work is Plan Mode — no files are modified.

Key Feature: 3-Tier Planning Pipeline

Deep (/athanor:deep-plan):
         ┌── Planner A ──→ Reviewer B ──┐
Input ───┤                               ├── Critic → Final Plan
         └── Planner B ──→ Reviewer A ──┘

Standard (/athanor:plan, default):
Input ── Planner ──→ Reviewer ──→ Refinement → Final Plan

Lite (/athanor:lite-plan):
Input ── Planner ──→ Final Plan
Tier When to use Cost
lite Quick iteration, low-risk changes 1 worker
standard Default for most tasks 2-3 workers
deep High-stakes architecture, complex refactors 5-6 workers

Design Philosophy

8 commands, not 100 features. A focused workflow instead of a feature collection. Each command maps to one phase: brainstorm, analyze, plan, execute.

Thin leader. The main session never reads files, analyzes code, or writes code. It dispatches and collects. Your context stays clean regardless of session length.

3-tier adversarial planning. Choose the review depth that fits the task: lite for speed, standard for balanced quality, deep for competing perspectives with cross-model review.

Grow with use. Lessons from every session are extracted, scored, and surfaced to future workers. The system gets better at your codebase over time.

Plan before execute. No files are modified until you confirm the plan. Explicit mode separation prevents accidental changes.

Session isolation. If a work-log exists in a session, a new session is automatically created. Every planning cycle starts clean.

Architecture

Thin leader dispatches to clean-context workers:

Agent Model Role
researcher sonnet Objective research + Devil's Advocate
analyst sonnet Fast parallel analysis
planner opus Implementation planning
critic opus Plan synthesis and review
executor opus Code execution with verification loop
learner sonnet Session learning extraction
task splitter sonnet Plan → subtask decomposition
cleaner haiku Memory decay and cleanup

When Codex is available, it serves as Planner B (deep tier) or Reviewer (standard tier).

Session communication via .md files — workers read and write to .athanor/sessions/{id}/. No shared state in the leader's context.

Learning system — after each /athanor:work, the Learner extracts structured lessons to .athanor/lessons/. Workers read relevant lessons before starting. Frequently-accessed lessons auto-promote to permanent. Stale ones decay and get cleaned.

Full architecture details | Conventions

Execution Modes

Solo (--solo) — One subtask at a time, each in a clean context. Simple and reliable.

Team (--team) — Wave-based parallel execution. Subtasks grouped by dependency, each wave runs simultaneously with discovery relay between waves.

Wave 1: [task 1, task 2]  ← parallel
  ↓ discovery relay
Wave 2: [task 3]          ← depends on wave 1

Configuration

athanor.json at project root (auto-created by /athanor:setup):

Key Default Description
codex.enabled true Cross-model planning with Codex (falls back to contrarian Claude)
work.defaultMode "solo" Default execution mode
work.ralphLoop.maxRetries 5 Max verification retries per subtask
work.circuitBreaker.consecutiveFailures 3 Failures before circuit breaker trips
team.waveSize 3 Max parallel workers per wave
memory.decayDays 7 Working memory retention period
memory.promotionThreshold 5 Access count for auto-promotion to permanent

FAQ

How is this different from just using Claude Code? Claude Code is one brain doing everything. Athanor gives it a team — separate workers for research, planning, execution, and review, each with clean context. The main session stays lightweight no matter how long you work.

Why cross-model adversarial planning?

Different models have different blind spots. In deep mode, Claude and Codex create competing plans and review each other's work. Standard mode gets Codex review on a single plan. Lite mode skips review entirely. Each tier degrades gracefully:

  • Deep: Planner B falls back to Claude contrarian; Reviewer B falls back to Claude.
  • Standard: Codex review falls back to Claude self-review.
  • Lite: Unaffected (Claude only by design).

What does /athanor:debug do?

Triage-first failure diagnosis. It analyzes errors, git history, and code traces in parallel to pinpoint root causes before you plan a fix.

Does this work without Codex? Yes. All tiers degrade gracefully — deep mode uses Claude contrarian instead of Codex as Planner B, standard mode falls back to Claude self-review. Set codex.enabled: true to use Codex when available.

How much token overhead does this add? The leader session stays minimal — it only dispatches and collects. Workers use tokens in clean contexts that are discarded after each task. Net cost is comparable to doing the same work manually, but with better plan quality.

Who is this for? Solo developers who want structured planning. Tech leads who want reproducible quality. Teams who want their Claude Code sessions to share learned lessons.

Companion Plugins

Athanor recommends (does not require) the following companion plugin:

superpowers — provides foundational skills like verification-before-completion. Athanor vendors a copy of this skill for guaranteed Stop hook behavior, but having superpowers installed adds the rest of its skill catalog (TDD, debugging, collaboration patterns).

# Official marketplace
/plugin install superpowers@claude-plugins-official

# Fallback marketplace
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Run /athanor:setup to audit installed companions. If superpowers is absent, athanor remains fully functional — its vendored skill ensures the Stop hook works regardless. See docs/DEPENDENCIES.md for the full policy.

Roadmap

  • Core workflow (discuss, analyze, plan, work)
  • Cross-model adversarial planning
  • Solo and team execution modes
  • 2-tier learning system with memory decay
  • Codex native integration (codex exec CLI)
  • Multi-project lesson sharing
  • Custom worker agent definitions
  • CI/CD integration hooks

Contributing

See CONTRIBUTING.md.

Docs

Acknowledgments

Athanor vendors the following open-source component:

  • verification-before-completion skill from superpowers by Jesse Vincent (MIT). Triggered by the athanor Stop hook to enforce evidence-before-claims discipline at turn boundaries. Full attribution in NOTICE.md.
  • scope-drift skill from claude-octopus by nyldn (MIT). On-demand skill that detects scope drift between current branch changes and the canonical plan-of-record. Full attribution in NOTICE.md.

License

MIT

About

Athanor — Self-sustaining agentic workflow orchestrator. Cross-model adversarial planning with thin-leader dispatch.

Resources

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages