Pattern-based scaffold for systematic AI agent work, with honest capability boundaries. Inspired by observed Claude Mythos Preview behaviors.
SaferSkills independently audited claude-mythos-scaffold (Agent Skill) and scored it 96/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 1 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 1 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Agentic discipline framework anchored by working patterns distilled from Claude Fable 5 itself (2026-07-06, revised 2026-07-21). v2 adds a persistent mission file, harness-enforced persistence (Stop/SessionStart hooks), named worker agents with explicit model routing, a headless profile, and an eval harness that measures the scaffold instead of trusting it.
Honest expectation: the scaffold transfers process discipline, not capability. It makes Opus 4.8 a disciplined Opus 4.8. See eval/ for measured deltas.
| Session model | What to load |
|---|---|
| Opus 4.8 / Opus 4.x / Sonnet 4.x / Haiku | Kernel, then tiered (below) |
| Claude 5 family (Fable 5 / Mythos 5 / Opus 5 / Sonnet 5) and later | Nothing, or at most the kernel's verification section. State goal and constraints; do not enumerate steps. |
When in doubt (a model newer than this table): treat it as Claude 5 family — under-scaffolding costs less than over-scaffolding on capable models.
Always load exactly one file: the kernel, core/fable-distilled.md. It carries the binding rules and the load map for every other file (mode, tool-stack, context-priming, decomposition, agent-loop, verification, failure-recovery, memory, headless, domains). Open others only when their load-map trigger fires.
Long-horizon runs externalize state to ~/.claude/mythos/missions/<slug>--<stamp>.md with an active-mission pointer at ~/.cache/mythos/active. The /mythos-mode command owns the lifecycle (<task> / resume / status / close); the hooks in hooks/ read it to block premature stops and to re-surface the mission after compaction. Schema: see the command's mission template.
agents/ ships four workers installed into ~/.claude/agents/: mythos-scout (haiku, mechanical bulk), mythos-builder (sonnet, light code), mythos-heavy (opus, hard code), mythos-verifier (sonnet, adversarial check, never edits). Every spawn names its model explicitly.
domains/research/ (retrieval tiers, claim graph,cite-verify, output formats)
domains/migration/ (audit, plan, execute, rollback)The scaffold routes to neighbors instead of duplicating them: see INTEGRATION.md.
core/headless.mdprotocol); mission file + pointer; Stop/SessionStart enforcement hooks; four named worker agents; headless profile; eval harness (T1 persistence, T2 verification) and distillation curation procedure; command grows resume/status/close; installed command becomes a symlink; stale references fixed; dead sync hook removed.
model gating; sub-agent model routing.
https://github.com/kasparovabi/claude-mythos-scaffold (MIT, by kasparovabi). Inspired by Claude Mythos Preview behaviors, not affiliated with Anthropic.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.