Architecture

System Layers

graph TB
    subgraph "Scheduling Layer"
        SCH[noether-scheduler]
    end

    subgraph "Composition Layer (Noether Engine)"
        TICK[sprint_tick.json]
        KICK[kickoff.json]
        RETRO[retro.json]
        SPAWN[spawn_agent.json]
    end

    subgraph "Stage Layer (Python)"
        GH[GitHub stages]
        DAG[DAG stages]
        SUP[Supervisor stages]
        RET[Retro stages]
        KF[Kickoff stages]
    end

    subgraph "Infrastructure Layer (Rust)"
        SHELL[caloron-shell ~200 LOC]
    end

    subgraph "Storage"
        KV[(Noether KV Store)]
        GITEA[GitHub / Gitea]
    end

    SCH --> TICK
    SCH --> KICK
    SCH --> RETRO
    TICK --> GH
    TICK --> DAG
    TICK --> SUP
    TICK --> SPAWN
    GH --> GITEA
    DAG --> KV
    SUP --> KV
    SHELL --> KV

What Lives Where

Concern	Original Caloron	Caloron-Noether
Orchestration loop	orchestrator.rs (404 lines)	sprint_tick.json + scheduler
State management	DaemonState in memory	Noether KV store (SQLite)
Event handling	git/monitor.rs (680 lines)	poll_events.py + dag_evaluate.py
GitHub API	git/client.rs (237 lines)	5 Python stages (~250 lines)
DAG engine	dag/engine.rs (627 lines)	evaluate.py (~130 lines)
Supervisor	4 Rust files (934 lines)	3 Python stages (~140 lines)
Retro	5 Rust files (2,083 lines)	3 Python stages (~170 lines)
Process management	agent/spawner.rs (336 lines)	shell/src/spawner.rs (~100 lines)
Heartbeat	Unix socket server (345 lines)	HTTP endpoint (~50 lines)

Data Flow: Sprint Tick

1. Scheduler triggers sprint_tick.json
2. Parallel: load DagState + last_poll + agents from KV
3. github_poll_events → fetch events since last_poll
4. Save new last_poll to KV
5. dag_evaluate → advance state machine, produce actions
6. Save updated DagState to KV
7. Parallel:
   a. check_agent_health → decide_intervention → supervisor actions
   b. dag_is_complete → check if sprint is done
8. execute_actions → dispatch: spawn agents, merge PRs, post comments

The Shell

The only Rust code. Three HTTP endpoints:

• POST /heartbeat — agents report liveness, written to KV
• POST /spawn — create git worktree + start harness process
• GET /status — list agent PIDs and liveness

The shell does no business logic — it only manages OS processes and proxies heartbeats to the KV store.