rustcircle/c101

A self-tuning, multi-agent stack for Claude Code and the Claude Agent SDK.
One delegator, orchestrators, swarms, meshes, expert ensembles and voting mediators — nested up to eight levels deep,
bounded by token budgets, and learning from every session.
## What it is
c101 is a complete Claude Code configuration (agents, skills, saved workflows, hooks, MCP servers and Python
tooling) that turns a single session into a coordinated team:
Orchestration. A main-thread delegator routes every request to the cheapest adequate route: a single expert,
an orchestrator running a plan DAG, a swarm over shards, a subagent mesh on a shared blackboard, an ensemble of
experts, or a mediator where several agents solve the same problem independently and vote (Schulze, Borda,
approval, plurality, logarithmic opinion pool).
63 agents in nine groups (orchestration, planning, quality, research, experts, synthesis, image, writing, meta), each
with a model tier, an effort level, a step budget, only the MCP servers it calls (connected only while it runs), and at most three role skills on
top of the three protocol skills every agent preloads.
Depth 8, bounded. Native nested subagents are raised from 3 to 8 levels. The pre-agent hook gives every child
a header (depth, parent, remaining depth, spawn count, token budget, output directory) and
enforces per-caller spawn rules, a 128-spawn session budget and whole-tree token budgets (session 2.5B guaranteed,
7.5B soft, 10B hard; per prompt 100M, 300M, 400M).
Three model tiers. claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5 are the only models allowed
anywhere, enforced at seven layers; a caller that wants cheaper or stronger work picks a different agent, never a
per-call model or effort.
Planning and quality. Planners, an architect, a plan critic, reviewers with adversarial verification,
validators, testers, quality control, code-quality analysis, optimisation, security and debugging; nothing reaches
the user without a gate.
Research. Quick and deep researchers, web search (Exa, Jina), scraping and crawling (Spider), real browsing
(Playwright), computer use (the user's own macOS apps interactively, or a disposable, network-less desktop
container that also runs under the Agent SDK), papers (arXiv/SSRN), fact checking, and summarizers, including a file-summarizer
that keeps full output on disk and returns a compact digest.
Experts and synthesis. Computer graphics, computer science, quantum computing, Rust, C++, Go, Haskell, Julia,
Java, Kotlin, Python, TypeScript, mathematics, physics and graphic design, each backed by a hub skill and granular skills, plus
synthesis agents that combine experts into higher-level answers (Delphi panels, debate, analogy, first principles).
Saved workflows. Nine deterministic JavaScript workflows (/swarm-run, /fanout, /mediate,
/expert-panel, /research-sweep, /review-verify, /plan-tournament, /visual-iterate, /manuscript-review): agents produce the
content, and the script computes every decision (tallies, quantiles, stop rules, dedup, coverage ledgers).
Context first. Output past ~60 lines goes to .stack/out/ / and only a summary and a path travel up;