Architecture

The tier-3 model

The goatcs-harness runtime is classified as tier-3 in a harness maturity model:

Tier-3 means the LLM is not merely assisted by context engineering but is programmed by a graph: the harness reads graph.json, computes the ready-set at each step, dispatches nodes, evaluates gate conditions, runs no-llm nodes as code, and checkpoints after each node.

Inline drivability

Skills can be driven in three modes:

• --provider inline — the orchestrator (you) reasons at each node. Deterministic nodes run as code; reasoning nodes pause at exit-11 for you to submit output, then resume. This is the primary mode for the epiphany pipeline.
• --provider claude-cli — the harness spawns claude -p for each reasoning node.
• --provider codex — the harness spawns Codex for each reasoning node.

The graph is identical across providers; only the live author differs.

Pipeline

The harness pipeline for a single run:

1. Load — read graph.json, validate against graph.schema.json (PRC-1).
2. Seed — load the seed file (initial state / inputs).
3. Ready-set loop — compute which nodes are ready (all predecessors complete, scale gates satisfied, branch labels active), dispatch them, collect output, checkpoint.
4. Gate evaluation — for forward-conditional edges, evaluate the gate_condition expression against the current signal state.
5. Verification — no-llm verifier nodes run deterministic checks.
6. Emit — atomic write of final artifacts (WAL + shadow-write + atomic rename).

Module map