Claude Mythos Scaffold
Pattern-based scaffold for systematic AI agent work, with honest capability boundaries. Inspired by observed Claude Mythos Preview behaviors (Anthropic, April 2026). Not affiliated with Anthropic.
A Claude Code skill set that brings agentic discipline (Plan-and-Execute, Reflexion, Self-Refine, Ralph Loop) into structured, reusable protocols. Since v0.3 the discipline is not only prose: a persistent mission file externalizes task state, Stop/SessionStart hooks enforce persistence at the harness level, named sub-agents carry explicit model routing, and a small eval harness measures the scaffold instead of trusting it.
Why This Exists
Claude Mythos Preview demonstrated agentic behaviors (problem persistence, multi-step iteration, verification-driven correction) that scaffold patterns can partially reproduce in Opus-class and smaller models (Opus 4.8, Sonnet 5, Haiku 4.5). This repo packages those patterns as opt-in skills. The kernel, core/fable-distilled.md, carries working patterns written by Claude Fable 5 (the GA'd Mythos-class model) about its own decomposition, verification, and next-action habits: first captured 2026-07-06, revised live by a Fable session 2026-07-21.
Honest expectation: the scaffold transfers process discipline, not capability. It makes Opus 4.8 a disciplined Opus 4.8, not a Fable. Raw reasoning depth, novel pattern recognition, and sample efficiency cannot be patched in via prompts; those are weights. Do not fully load it on Fable 5 / Mythos 5: over-scaffolding degrades Mythos-class output.
What you get: a compact always-load kernel, demand-loaded protocols for tool selection, priming, decomposition, verification and failure recovery, a mission-file lifecycle with enforcement hooks, four worker agent definitions, a headless profile for unattended runs, an eval harness, and two domain modes (research synthesis, codebase migration).
Repository Layout
claude-mythos-scaffold/
├── SKILL.md Skill entry: model gating + kernel activation
├── INTEGRATION.md Route-don't-duplicate table (neighbor skills)
├── core/
│ ├── fable-distilled.md The kernel: Fable 5 patterns + binding rules (read first)
│ ├── mode.md Framing, honest ceiling, operating template
│ ├── tool-stack.md Cascade selection, parallel/sequential
│ ├── context-priming.md Adaptive RAG, source hierarchy
│ ├── decomposition.md Sub-agent delegation, model routing
│ ├── agent-loop.md Plan-Execute, Reflexion, Self-Refine
│ ├── verification.md Headless verify, output reading
│ ├── failure-recovery.md Ralph Loop, persistence threshold
│ ├── memory.md Cross-session memory integration
│ └── headless.md Preamble for unattended runs (cron, loops, eval)
├── agents/ Worker definitions (install into ~/.claude/agents/)
│ ├── mythos-scout.md haiku: mechanical bulk, references only
│ ├── mythos-builder.md sonnet: scoped light implementation
│ ├── mythos-heavy.md opus: hard multi-file work
│ └── mythos-verifier.md sonnet: adversarial verification, never edits
├── hooks/
│ ├── mythos-stop.py Stop hook: blocks premature stops (bounded nudges)
│ ├── mythos-session.py SessionStart hook: mission survives compaction
│ ├── autoskill-mythos-rule.snippet Optional prompt-router rule (documented)
│ ├── install.sh Copies hooks, prints settings JSON
│ └── README.md Contract, registration, kill-switch
├── eval/
│ ├── mythos-eval Runner: bare vs scaffold arms, headless
│ ├── tools/metrics.py Objective metrics from stream-json transcripts
│ ├── tasks/T1-persistence/ Planted-obstacle fixture + check.sh
│ ├── tasks/T2-verification/ Tempting-wrong-fix fixture + check.sh
│ └── distill/CURATION.md Distillation curation procedure (manual)
├── domains/
│ ├── research/ mode, retrieval, synthesis, cite-verify, output
│ └── migration/ mode, audit, plan, execute, rollback
└── commands/
└── mythos-mode.md /mythos-mode: mission lifecycle + full pass
Runtime state lives outside the repo: missions in ~/.claude/mythos/missions/, the active-mission pointer in ~/.cache/mythos/active, eval data in ~/.claude/mythos/eval-data/.
Quick Start
Option A: Symlink into Claude Code (recommended)
macOS or Linux:
git clone https://github.com/kasparovabi/claude-mythos-scaffold ~/.claude/skills/mythos-scaffold
cd ~/.claude/skills/mythos-scaffold
ln -sf "$PWD/commands/mythos-mode.md" ~/.claude/commands/mythos-mode.md
mkdir -p ~/.claude/agents
for f in agents/mythos-*.md; do ln -sf "$PWD/$f" ~/.claude/agents/"$(basename "$f")"; done
bash hooks/install.sh # optional: enforcement hooks; prints the settings.json entries
Symlinks keep the installed command and agents in lockstep with the repo. If your platform lacks symlink support, cp works identically (the command uses absolute paths, no macro resolution).
Then in any Claude Code session:
/mythos-mode <your task>
/mythos-mode resume | status | close
Option B: Domain-only
Just need research synthesis or codebase migration? Copy the relevant domains/<name>/ directory plus core/ (the kernel is required).
Option C: Vault integration (Obsidian-style)
If you maintain a knowledge vault, the skills support [[wikilink]] cross-references and integrate with MemPalace for verbatim conversation history. See core/memory.md.
What's Tested vs Newer
Production-tested (used across multiple real Claude Code sessions):
core/mode.md,tool-stack.md,context-priming.md,decomposition.md,agent-loop.md,verification.md,failure-recovery.md
Newer (v0.3, self-tested + eval-smoked, real-world feedback expected):
core/fable-distilled.mdkernel revision, mission-file lifecycle,hooks/*,agents/*,core/headless.md,eval/*core/memory.md,domains/research/*,domains/migration/*
Acknowledgments
This scaffold is a synthesis of patterns from active open-source projects, research papers, and industry case studies. Full credit list with links: see REFERENCES.md.
Highlights:
- Claude Mythos Preview (Anthropic, April 2026): the observed behaviors that motivated this scaffold
- MemPalace (Jovovich + Sigman, 2026): AAAK compression, hierarchical loci memory, used in
core/memory.md - claude-mem (thedotmack): Claude Code-native memory plugin alternative
- Ralph Loop (Geoffrey Huntley): persistence pattern, used in
core/failure-recovery.md - Self-Refine (Madaan et al.): iterative output refinement, used in
core/agent-loop.md - Reflexion (Shinn et al.): episodic self-critique
- ReAct (Yao et al.): think-act-observe loop
- Adaptive RAG (Jeong et al.): query-complexity routing, used in
domains/research/retrieval.md - PaperOrchestra (Google Research): multi-agent research synthesis
- Environment-in-the-Loop (ReCode 2026): code migration with environment integration
- Aviator Java to TypeScript case study, Doctolib production migration (2026 enterprise reports)
This repo is independent. Not endorsed by Anthropic, MemPalace, or any cited project. Patterns are credited; opinions and integration choices are mine.
Contributing
PRs welcome, see CONTRIBUTING.md. Particular areas of interest:
- Real-world session case studies (anonymized) for an
examples/directory - More eval tasks (decomposition, orchestration axes) and cross-model baselines
- Generic versions of vault-specific references (some skills mention conventions like
(C)prefix; these are opt-in) - Additional domain modes (data engineering, devops incident response, content writing)
License
MIT, see LICENSE.
Anti-Goals (What This Is Not)
- "10x your AI agent" magic. Scaffolding helps consistency, not raw capability.
- Drop-in replacement for engineering judgment. The scaffold structures decisions; humans still make them.
- Production-stable v1.0. This is v0.3; patterns are sound, edge cases will surface.
If you're looking for those things, this repo will disappoint. If you're looking for systematic AI agent discipline with honest documentation, it's a starting point.
No comments yet
Be the first to share your take.