sancovp/application/compoctopus/skills/chain-architecture

Chain Architecture

You design agent chains for the Compoctopus pipeline.

Chain Types

Chain (sequential)

Links execute in order. Output of link N feeds into link N+1.

Link1 → Link2 → Link3 → result

Use when: each step has a clear input/output, no retry logic needed.

EvalChain (loop with evaluator)

Chain runs, then an evaluator approves/rejects/retries.

Link1 → Link2 → Link3 → Evaluator
                              ↓
                   approve / reject / retry
                              ↓ (retry)
                   Link1 → Link2 → ...

Use when: output quality matters, annealing/iteration improves results.

Link Types

| Type | When to use | Has LLM? | |------|-------------|----------| | SDNAC | Agent reasoning, generation, analysis | Yes | | FunctionLink | Data transforms, validation, file I/O | No |

Design Patterns

The Annealer (OctoCoder pattern)

STUB → TESTS → PSEUDO → ANNEAL → VERIFY

Each phase is an SDNAC. VERIFY evaluates. On failure, restarts with existing work.

The Decomposer (Planner pattern)

PROJECT → FEATURES → COMPONENTS → DELIVERABLES → TASKS

Sequential Chain. Each phase adds one level of GIINT hierarchy.

The Router (Bandit pattern)

SELECT (try golden chain) → if miss → CONSTRUCT (build new chain)

EvalChain. Exploit known solutions before exploring new ones.

Dovetail Rules

• Every link MUST declare inputs and outputs
• Output names of link N must match input names of link N+1
• Type mismatches fail at chain construction time, not runtime