Coding Agent (Idle Compute)

Dispatches coding tasks to the CREATE/REUSE pipeline during node idle time.

How It Works

When a node detects idle compute capacity, the coding agent picks up pending coding goals from the network and executes them through the standard CREATE/REUSE pipeline. This turns unused compute into productive work for the network.

Tool Backends

The orchestrator selects the best backend for each task:

Backend	License	Best For
KiloCode	Apache 2.0	General-purpose code generation; open-source friendly.
Claude Code	Proprietary	Complex reasoning and multi-file refactoring.
OpenCode	MIT	Lightweight tasks; fully open-source.

The orchestrator considers task complexity, required context window, and the node's compute policy when choosing a backend. If the compute policy is local_only, only backends that can run locally are eligible.

Task Flow

1. Idle detection -- The node's resource monitor detects idle CPU/GPU capacity.
2. Task pickup -- The coding agent queries for pending coding goals.
3. Backend selection -- The orchestrator picks the best tool backend.
4. Execution -- The task is dispatched to the CREATE/REUSE pipeline.
5. Result delivery -- Output is written back to the goal's ledger entry.

Metered API Costs

If the coding agent uses a metered backend (e.g., Claude Code) for a hive task, the cost is recorded in MeteredAPIUsage with task_source=idle and compensated via settle_metered_api_costs(). See metered-api-recovery.md.

Source Files

• integrations/coding_agent/
• integrations/agent_engine/speculative_dispatcher.py