Improve architecture

Iteration loop: explore for friction, present deepening candidates, grill the chosen one, update domain artifacts inline. Vocabulary is load-bearing. See references/LANGUAGE.md.

Vocabulary [LOAD-BEARING]

Use these terms exactly. Do not substitute "component," "service," "API," or "boundary." Full definitions in references/LANGUAGE.md.

• Module: anything with an interface and an implementation (function, class, package, slice). Scale-agnostic.
• Interface: every fact a caller must know: types, invariants, ordering, error modes, config, performance shape. Not just signature.
• Depth: leverage at the interface. Deep = much behaviour behind a small interface. Shallow = interface as complex as implementation.
• Seam (Feathers): where an interface lives; a place behaviour can be altered without editing in place. Use this, not "boundary."
• Adapter: a concrete thing satisfying an interface at a seam. Role, not substance.
• Leverage: capability callers gain per unit of interface learned.
• Locality: concentration of change, bug, and knowledge at one site for maintainers.

Principles

• Deletion test: imagine deleting the module. If complexity vanishes, it was a pass-through. If complexity reappears across N callers, it was earning its keep.
• Interface = test surface. Tests cross the same seam callers cross. Wanting to test past the interface = wrong shape.
• One adapter = hypothetical seam. Two adapters = real seam. No port without two real implementations (production + test).

Process

1. Explore

First tool call MUST be Explore-agent dispatch, not direct reads. The agent's brief:

• Read CONTEXT.md (or CONTEXT-MAP.md + per-context CONTEXT.md) and any docs/adr/. If absent, proceed silently.
• Walk the codebase organically; classify friction:
  • Concept understanding requires bouncing across many small modules → shallow cluster.
  • Interface complexity matches implementation complexity → shallow module.
  • Pure functions extracted only for testability while real bugs hide in callers → no locality.
  • Tightly coupled modules leaking across their seams.
• For each suspect, run the deletion test before reporting.

2. Present candidates

Numbered list. Each candidate: Files, Problem (concrete friction; cite deletion test), Solution (plain-English description; no interface yet), Benefits (locality, leverage, testability deltas).

ADR conflicts: surface only when friction warrants reopening; mark explicitly: "contradicts ADR-0007: worth reopening because…".

Ask: "Which candidate to explore?" Do not propose interfaces yet.

3. Grilling loop

Once user picks, drop into adversarial interview. Walk the design tree, resolve dependencies one decision at a time, recommend an answer per question. Side effects happen inline:

• New domain term emerging? Update CONTEXT.md immediately (lazy create).
• User rejects with a load-bearing reason that future explorers would need? Offer ADR.
• User wants alternative interfaces for the chosen candidate? Pivot to references/INTERFACE-DESIGN.md for parallel sub-agent design twice (Ousterhout).

Deepening categories (testing strategy per dependency class)

Full treatment in references/DEEPENING.md. Summary:

Class	Deepenable?	Test strategy
In-process (pure / in-memory)	Always	Merge modules; test through new interface directly. No adapter.
Local-substitutable (PGLite, in-memory FS)	Yes if stand-in exists	Stand-in runs in tests; seam stays internal.
Remote but owned (microservices)	Yes via Ports & Adapters	Port at seam; HTTP/gRPC adapter prod, in-memory adapter test.
True external (Stripe, Twilio)	Yes	Injected port; mock adapter for tests.

Replace, don't layer: delete shallow-module tests once interface tests exist.

Language-neutral examples

Rust: shallow validate_address + format_address + geocode_address separately called by a Shipment aggregator. Deletion test: removing format_address concentrates string-handling at one call site → was a pass-through. Deepen into address::Resolver with resolve(raw) -> Result<Resolved, AddressError>; tests cross the new interface; in-memory Geocoder adapter for tests, HTTP adapter for production.

Python: module exposes parse_invoice, apply_tax, round_total as separate top-level functions; every caller chains all three. Deepen into billing.Invoice.finalize(raw) -> Invoice. Internal seams (tax tables, rounding rules) stay private; the test surface is Invoice.finalize.

Reference docs

• references/LANGUAGE.md: full vocabulary, principles, rejected framings.
• references/DEEPENING.md: dependency taxonomy, seam discipline, replace-don't-layer testing.
• references/INTERFACE-DESIGN.md: parallel sub-agent "Design It Twice" workflow when the chosen candidate's interface needs alternatives.

Forbidden: proposing interfaces in step 2 (premature commitment), bundling unrelated refactors, re-litigating ADRs without a load-bearing reason.