improve-architecture

Purpose

Surface architectural friction in a codebase and propose deepening opportunities — refactors that turn shallow modules into deep ones. Output is structured (numeric scores + JSON proposal payloads) so surgeon, review, and audit can consume it programmatically without re-reading the codebase.

The goal is testability and AI-navigability: a deep module presents a small interface that hides large machinery, so tests target one surface and future agents can reason about the system without traversing N small wrappers.

Vocabulary (controlled — use exactly)

These eight terms have precise meanings. Banned aliases: "boundary" (overloaded with DDD), "component" (UI-specific), "service" (microservice-specific), "layer" (too generic). See references/language.md for full definitions.

• Module — anything with an interface and an implementation (function, class, package, slice).
• Interface — everything a caller must know to use the module: types, invariants, ordering, error modes, config.
• Implementation — the code inside.
• Depth — leverage at the interface; large behavior behind a small interface.
• Seam — where an interface lives; place behavior can be altered without editing in place.
• Adapter — concrete thing satisfying an interface at a seam.
• Leverage — what callers get from depth.
• Locality — what maintainers get from depth.

Triggers

• Called by cook Phase 5 (quality gate) when refactor signals appear in scout output
• Called by surgeon before any deepening session — produces the proposal surgeon executes
• Called by audit to compute the architecture sub-score
• Called by review when a reviewer flag mentions "shallow", "wrapper", "indirection"
• Manual: /rune improve-architecture <module-path>

Calls (outbound)

• scout (L2): re-scan target module + callers when input context is stale
• brainstorm (L2): when the deepened module needs a new interface, hand off in design-it-twice mode (see brainstorm v0.6+)
• journal (L3): record an ADR if the user rejects a candidate with a load-bearing reason

Called By (inbound)

• cook (L1): Phase 5 quality gate
• surgeon (L2): pre-refactor input; consumes the proposal payload
• audit (L2): Phase 4 architecture sub-score
• review (L2): when shallow-module flag fires during review
• User: manual invocation

Cross-Hub Connections

• improve-architecture → surgeon — proposal payload feeds surgeon's deepening session
• improve-architecture ↔ brainstorm — when interface needs design-it-twice exploration
• improve-architecture → audit — emits architecture sub-score
• improve-architecture → journal — records ADRs for rejected candidates with load-bearing reasons

Inputs

• Required: target module path (e.g. src/auth/) OR signal codebase.scanned from a recent scout pass
• Optional: existing CONTEXT.md (domain glossary, used to name modules in their domain language)
• Optional: docs/adr/ directory (existing ADRs that constrain proposals — do not re-litigate)

Executable Steps

Step 1 — Read existing context

Read in order, silently skipping any that don't exist:

1. CONTEXT.md (or CONTEXT-MAP.md + per-bounded-context CONTEXT.md)
2. Relevant docs/adr/ files
3. The target module's source files (use Glob to enumerate, cap at 30 files)
4. Direct callers of the module (grep for imports / require / use)

If CONTEXT.md is missing, do not flag it — treat as "no domain glossary yet". If an ADR contradicts a candidate you're forming, mark it and only surface the candidate if the friction is genuine enough to warrant ADR revision.

Step 2 — Score the candidate(s)

For each candidate module, compute three numeric scores (1–5) and one verdict (enum):

Metric	Formula / Rubric
Depth	clamp_1_5(implementation_complexity / interface_complexity) — 1 = shallow wrapper, 5 = small interface hides large machine
Leverage	clamp_1_5(num_callers * unique_use_cases / interface_method_count) — 1 = thin caller benefit, 5 = many callers, fewer methods to learn
Locality	clamp_1_5(code_concentration_index) — 1 = logic spread across N callers, 5 = logic concentrated in one place
Deletion test	enum: vanish (was pass-through) | concentrate (was earning keep) | redistribute (mixed)

Rubric details and edge cases: see references/scoring.md.

Step 3 — Classify dependencies

For each candidate's external dependencies, classify into one of four categories. The category determines test strategy:

Category	Definition	Test Strategy
in-process	Pure computation, in-memory state, no I/O	Test through deepened interface directly
local-substitutable	Has local test stand-in (PGLite, in-memory FS)	Use stand-in in tests; no port at module seam
remote-owned	Your own module deployed across a network seam	Define a port; in-memory adapter for tests, HTTP adapter for prod
true-external	Third-party (Stripe, Twilio)	Inject as port; mock adapter in tests

Full doctrine in references/deepening.md.

Step 4 — Apply seam discipline

Before recommending a port:

• One adapter = hypothetical seam. Two adapters = real seam. Don't introduce a port unless ≥2 adapters are justified (typically prod + test).
• Single-adapter "seams" are flagged "indirection-only" and dropped from the proposal.
• Internal seams (private to the implementation) MAY exist for the deepened module's own tests; they don't appear in the public interface.

Step 5 — Emit proposal payload

For each surviving candidate, produce a structured proposal in YAML:

architecture.proposal:
  module_path: src/auth/
  current:
    depth: 2
    leverage: 3
    locality: 2
    deletion_test: redistribute
  target:
    depth: 4
    leverage: 4
    locality: 4
  dependency_category: remote-owned
  suggested_seam: AuthPort
  adapters_planned: [HttpAuthAdapter, InMemoryAuthAdapter]   # 2 = real seam ✅
  tests_to_replace: [auth/login.test.ts, auth/session.test.ts]
  tests_to_write_new: [auth/AuthPort.test.ts]
  domain_terms_used: [Customer, Session]   # from CONTEXT.md if present
  adr_conflicts: []

Step 6 — Present candidates to user

Numbered list, each candidate showing:

• Files involved — file paths (+ key types/exports for durability)
• Problem — friction in the current architecture, in vocab terms (depth/leverage/locality)
• Solution — plain English, naming the deepened module by its domain term if CONTEXT.md provides one
• Benefits — leverage gain (caller-side) + locality gain (maintainer-side) + test surface change
• Score delta — current → target

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

Step 7 — On user pick

When user picks a candidate, hand off:

• To brainstorm in design-it-twice mode if the new interface is non-obvious (multiple credible shapes)
• To surgeon with the proposal payload otherwise

If user rejects a candidate with a load-bearing reason ("we don't want to centralize auth because of compliance audit isolation"), offer to record an ADR via journal (only if score >= 11 per journal v0.4 criteria).

Output Format

## Architecture Improvement Report

### Target
- **Path**: src/auth/
- **CONTEXT.md present**: yes / no
- **ADRs reviewed**: 3 (none conflicting)

### Candidates

#### 1. Auth port consolidation (depth 2 → 4)
- **Files**: src/auth/login.ts, src/auth/session.ts, src/auth/middleware.ts
- **Problem**: 3 shallow modules each handle one HTTP-flavored verb; logic about `Customer` identity is split across all three (locality = 2)
- **Solution**: collapse into AuthPort exposing `authenticate`, `revoke`, `verify` — 3 methods, deep impl
- **Benefits**: callers learn 3 methods instead of N free functions; auth logic concentrated; tests target the port
- **Score delta**: depth 2→4, leverage 3→4, locality 2→4
- **Deletion test**: redistribute (current modules ARE doing work, just spread)

#### 2. ...

### Recommendation
Candidate 1 — strongest leverage gain. Hand off to `brainstorm` design-it-twice for the AuthPort shape (3 credible alternatives), then `surgeon`.

### Architecture sub-score
- Current: 58/100
- Projected after candidate 1: 78/100

Returns

Artifact	Format	Location
Architecture Improvement Report	Markdown	inline
Proposal payloads	YAML	inline (per candidate)
Architecture sub-score	integer 0-100	inline + emitted to audit
ADR draft (if user rejects with load-bearing reason)	Markdown	.rune/adr/ADR-NNN-<slug>-s<score>.md via journal

Constraints

1. MUST use the 8 controlled vocabulary terms exactly — no aliases ("boundary", "component", "service", "layer" are banned in skill output)
2. MUST include numeric scores (depth/leverage/locality 1-5 each) on every candidate — soft prose claims are rejected
3. MUST apply deletion test verdict — "vanish" candidates may be removed entirely; "concentrate" candidates are deepening targets
4. MUST apply two-adapter rule — single-adapter seams are flagged "indirection-only" and dropped
5. MUST NOT propose interfaces in the same step as candidate selection — present candidates first, hand to brainstorm Design-It-Twice if interface is non-obvious
6. MUST silently skip missing CONTEXT.md / ADR directory — do not flag as project gap
7. MUST emit JSON-shaped proposal payload — downstream skills (surgeon) consume it programmatically

Sharp Edges

Failure Mode	Severity	Mitigation
Recommending a deepening that contradicts a documented ADR	HIGH	Step 1 reads ADRs; if conflict, surface only if friction is real enough to revise the ADR
Single-adapter seam slips into proposal	HIGH	Step 4 rule — drop or downgrade to "internal seam"
Vocabulary drift (using "boundary"/"component"/"service")	MEDIUM	Constraint 1 + linter pass in compiler/tests/vocabulary-discipline.test.js
Score inflation to make weak candidate look strong	HIGH	Each metric has rubric in scoring.md; judges show formula inputs
Missing CONTEXT.md domain terms — generic naming ("AuthService")	MEDIUM	If CONTEXT.md exists, names MUST come from it; otherwise OK
Proposing interface in same pass as candidates	MEDIUM	Step 6 hard-stops at candidate list; interface design = brainstorm Design-It-Twice
User rejects all candidates → no ADR recorded → next session re-litigates	MEDIUM	If reason is load-bearing AND score >= 11, offer journal ADR write

Self-Validation

SELF-VALIDATION (run before emitting Report):
- [ ] Every candidate has depth + leverage + locality scores (1-5 each)
- [ ] Every candidate has deletion-test verdict (vanish | concentrate | redistribute)
- [ ] Every candidate names a dependency category (in-process | local-substitutable | remote-owned | true-external)
- [ ] No banned vocabulary (grep candidate text for: boundary, component, service, layer in narrative)
- [ ] No interfaces drafted yet — that's brainstorm's job
- [ ] CONTEXT.md domain terms used if file present
- [ ] Each adapter list has >=2 entries OR seam is marked "internal-only"
IF ANY check fails → fix before reporting done.

Done When

• Target module read + callers mapped
• ≥1 candidate scored on all 3 axes + deletion test
• Proposal payload(s) emitted in valid YAML
• Architecture sub-score computed (0-100)
• User has either picked a candidate (handed to brainstorm/surgeon) or rejected with reason (ADR offered)
• Report emitted with vocabulary discipline intact

Cost Profile

~3000-7000 tokens input (codebase scan), ~2000-4000 tokens output (analysis + proposals). Opus model — architectural reasoning depth is the value. Called at most once per audit session, on-demand from cook / surgeon.

Chain Metadata

chain_metadata:
  skill: "rune:improve-architecture"
  version: "0.1.0"
  status: "[DONE]"
  domain: "[module path scored]"
  exports:
    architecture_subscore: 0-100
    candidates: [{ module, depth, leverage, locality, verdict }]
    proposal_payloads: [<yaml-per-candidate>]
  suggested_next:
    - skill: "rune:brainstorm"
      mode: "design-it-twice"
      reason: "Top candidate has multiple credible interface shapes — need diverse exploration before commit"
      consumes: ["proposal_payloads"]
    - skill: "rune:surgeon"
      reason: "User picked candidate; interface shape is obvious; ready for deepening session"
      consumes: ["proposal_payloads"]
    - skill: "rune:journal"
      reason: "User rejected candidate with load-bearing reason; record ADR (score >=11)"
      consumes: ["candidates", "rejection_reason"]

Scope guardrail: improve-architecture proposes and scores. It NEVER edits code. Refactor execution belongs to surgeon. Interface exploration belongs to brainstorm Design-It-Twice mode.