Caloron-Noether

Caloron reimplemented as Noether composition graphs. Same orchestration capabilities, ~6x less code.

Architecture

noether-scheduler (cron: sprint_tick.json every 60s)
        │
        ▼
Noether Engine (runs composition graphs)
  ├── DAG stages (Pure Python) — evaluate, is_complete, validate
  ├── GitHub stages (Network Python) — poll, create_issue, comment, merge
  ├── Supervisor stages (Pure Python) — health check, intervention, messaging
  ├── Retro stages (Pure/Network Python) — feedback, KPIs, report
  └── Kickoff stages (Network/LLM Python) — repo context, DAG generation
        │
        ▼
caloron-shell (~200 lines Rust, axum)
  ├── POST /heartbeat — record agent heartbeats
  ├── POST /spawn — create worktree + start harness
  └── GET /status — list live agents

All business logic is in Noether stages (Python). The shell only manages processes and HTTP.

Prerequisites

Caloron-Noether is a set of stages and compositions that run on top of Noether (v0.3.0+). You need both the noether CLI and noether-scheduler on your PATH.

cargo install noether-cli noether-scheduler
# or grab prebuilt binaries: https://github.com/alpibrusl/noether/releases/latest

See the Noether docs for deeper configuration, and the Scheduler guide for cron-driven compositions.

Setup

# 1. Verify noether is installed
noether --version
noether-scheduler --version

# 2. (Optional) Point at the hosted stage registry
export NOETHER_REGISTRY=https://registry.alpibru.com

# 3. Register custom stages
./register_stages.sh

# 4. Build the shell
cargo build -p caloron-shell

# 5. Start the shell (heartbeat + spawn server)
CALORON_SHELL_PORT=7710 ./target/debug/caloron-shell

# 6. Start the scheduler (drives sprint ticks + weekly retro)
noether-scheduler --config scheduler.json

Documentation

Full docs: docs/ — build locally with mkdocs serve.

• Getting Started
• Architecture
• Stage Catalog
• Composition Graphs
• Shell API
• KV Conventions
• vs Original Caloron
• Deployment (Docker + K8s)

Stage Promotion Path

All stages start as Python. When a stage meets all four criteria (generality, hot path, stable schema, worth the lines), it can be promoted to Rust via InlineRegistry — zero graph changes needed. See context/inline-stages.md.

CLI

Installing the package (pip install caloron-noether) exposes a caloron ACLI-compliant command:

caloron init my-project --backend noether            # create a project
caloron sprint "Build a hotel rate anomaly detector" # run an autonomous sprint
caloron status                                       # active project + last sprint
caloron history --limit 10                           # past sprints
caloron show 5                                       # full retro for sprint #5
caloron metrics --output json                        # aggregated KPIs
caloron agents                                       # agent profiles in this project
caloron projects list | switch | delete              # multi-project management
caloron config get|set <key> [value]                 # per-project settings

All commands accept --output text|json|table. The framework selected at caloron init --framework is propagated to the PO, HR, and reviewer agents. Supported frameworks: claude-code, cursor-cli, gemini-cli, codex-cli, open-code, aider. Non-claude frameworks use their agentic / auto-approval mode (-y for Gemini, --yes-always for Aider, exec --full-auto for Codex, -p for cursor-agent); make sure the corresponding CLI is authenticated and on $PATH.

Organisation conventions

Teams with a standard way of laying out projects, naming packages, or licensing files can declare those rules once and have every caloron sprint apply them. caloron org init scaffolds the file; edit it with your conventions:

caloron org init              # writes ~/.caloron/organisation.yml
# edit ~/.caloron/organisation.yml
caloron org show              # preview the block every agent will see
caloron org validate          # sanity-check the YAML

Per-project overrides go in <project>/caloron.yml (right-wins merge). Conventions are injected into the PO, agent, reviewer, and fix prompts — they're not enforced at tool level (no ruff config generation, no CI gate) today; the reviewer is the backstop.

Runtime dependency — Gitea

Caloron uses Gitea as the version-control backend for sprints — issues, PRs, merges, review comments all go through its API. A running Gitea container is required; without it, caloron sprint aborts with instructions. Start one with:

docker run -d --name gitea -p 3000:3000 -p 222:22 gitea/gitea:1.22

Set GITEA_TOKEN (or accept the dev-mode default) and configure your project with caloron config set repo <owner>/<repo> pointing at a repo you've created in that Gitea instance.

Bypass the preflight with caloron sprint --skip-gitea-check if you intentionally want to run without version control (agent still runs; all git/issue/PR calls become no-ops).

Sandbox

On Linux, caloron sprint runs each agent inside a bwrap (bubblewrap) sandbox. On macOS and other systems where bwrap is unavailable, it falls back to a no-op passthrough script so pip install caloron-alpibru works out of the box. Override with SANDBOX=/path/to/your-sandbox.sh if you need custom isolation.

Project Structure

orchestrator/         Sprint runtime (Python)
  orchestrator.py     Main loop: PO → agents → PRs → reviews → retro
  skill_store.py      Registry of skills/MCPs (18 built-in, user-extensible)
  hr_agent.py         Assigns skills + model + framework per task
  agent_configurator.py  Writes CLAUDE.md/.cursorrules/GEMINI.md + MCP configs
  agent_versioning.py    Tracks agent evolution across sprints
  template_store.py      Project scaffolds (YAML templates + LLM generation)
templates/            Project templates (user-extensible YAML)
  fastapi.yaml        FastAPI + ruff + pytest + Dockerfile
  fastapi-postgres.yaml  + SQLAlchemy + Alembic + docker-compose
  python-data.yaml    pandas + data dir + fixtures
  nextjs.yaml         Next.js 14 + TypeScript + Tailwind
  rust-cli.yaml       clap + clippy + fmt CI
stages/               Noether Python stages (stdin JSON → stdout JSON)
  dag/                DAG evaluation, completion, validation
  github/             GitHub/Gitea API operations
  supervisor/         Health checks, interventions, messaging
  retro/              Feedback, KPIs, report generation
  kickoff/            Repo context, DAG generation
compositions/         Noether composition graphs (JSON)
shell/                Thin Rust binary (axum HTTP server)
scripts/              Sandbox (bubblewrap on Linux, passthrough on macOS/other)
deploy/               Docker Compose + Kubernetes Helm chart
demo/                 Asciinema recording script