Architecture Audit

Overview

System-level health check for Reactive Agents. Like /simplify but scoped to architecture: surveys the package topology, kernel internals, and documentation for dead code, over-abstraction, layer violations, and stale claims — then fixes what's safe and flags what needs planning.

Use when:

• AGENTS.md, MEMORY.md, or inline docs may no longer match the code
• A package or subsystem has grown unexpectedly complex
• Cleaning up after a large feature: scaffolding, flags, disabled code wasn't removed
• Preparing to plan a refactor and need an honest current-state baseline

Don't use for:

• Effect-TS abstraction candidates specifically → effect-abstraction-audit
• Code-level cleanup of recent changes → /simplify
• Single-file or unit-level concerns

---

Phase 1: Architecture Snapshot

Authoritative current-state map: [[wiki/Architecture/Framework-Architecture-Index]] (file:line-anchored).

Before launching agents, orient to current state. Always start with the wiki — see [[wiki/Development/Wiki-Workflow|Wiki Workflow]].

claude-obsidian:wiki-query "<subsystem> architecture decisions"

This surfaces relevant: prior architecture audits, decisions, debriefs, and concept pages — avoiding rediscovery.

Read in this order:

Authoritative docs:

AGENTS.md                                        # canonical architecture + build order
wiki/Hot.md                                       # current empirical state
wiki/Architecture/Specs/09-UNIFIED-PROGRAM.md    # canonical sequencing + architecture target
wiki/Architecture/DEBT-REGISTER.md               # canonical technical debt
apps/cortex/AGENTS.md                            # Cortex-specific patterns

Wiki indexes:

wiki/Decisions/Decision Index.md                 # past architectural decisions
wiki/Issues/Running Issues Log.md                # known architectural debt
wiki/Failure-Modes/00 FM Catalog.md              # observed failure modes

Kernel internals (highest-churn area):

packages/reasoning/src/kernel/
  kernel-state.ts        # KernelState, Phase type, KernelContext
  kernel-runner.ts       # runKernel() loop
  react-kernel.ts        # makeKernel() factory
  phases/                # context-builder, think, guard, act
  utils/                 # ics-coordinator, loop-detector, tool-utils, etc.

Known architectural debt (pre-loaded, refreshed 2026-07-22 post-v0.14 audit — verify each is still present):

Debt item	Location	Status to verify
Provider adapter hooks (resolved v0.14: 4-hook system, all wired)	provider-adapters/	continuationHint, errorRecovery, synthesisPrompt, qualityCheck live (+ parseToolCalls); taskFraming/toolGuidance/systemPromptPatch removed
B3 builder→runtime seam structural cast	builder/build-effect/runtime-construction.ts	self as unknown as BuilderRuntimeStateView — renamed field NOT a compile error; Wave-4 codegen candidate
Loop-detector streak can mask duplicate-tool patterns	loop-detector.ts / kernel runner	strategySwitching may never trigger (W8) — still open?
PlanExecuteConfigSchema.patchStrategy dead config	reasoning/src/types/config.ts:32	declared + test-asserted, zero source readers
ToT outer loop ignores dispatcher-early-stop	plan-execute.ts / ToT sub-kernels	outer branches are separate sub-kernels
Sub-agent fixed vs dynamic path duplication	local-agent-tools.ts / spawn-handlers.ts	fixed path could be singletonAgentMode of dynamic

Resolved — do NOT resurface (verified 2026-07-22): context-engine.ts deleted entirely (with its flags); KernelState.meta is typed KernelMeta (0 as any in reasoning src); no disabled local-model strategy routing exists in strategy-registry.ts; define-tool DX (typed errors, Standard Schema, async handlers); local.ts timeout precedence; build() fail-fast on missing keys. Canonical ledger: wiki/Architecture/DEBT-REGISTER.md.

Determine scope before proceeding. Full repo, one package, or one subsystem? Narrow scope = deeper findings.

---

Phase 2: Launch Three Agents in Parallel

Use the Agent tool to launch all three concurrently. Pass each agent the relevant file paths and the debt table from Phase 1.

Agent 1: Architecture Health

Inspect the codebase for structural problems:

1. Layer violations — Does code in one layer reach through to another it shouldn't? Does act.ts know about provider specifics? Does context-builder.ts trigger side effects? Does reasoning package import directly from runtime?
2. Over-abstraction — Thin wrappers that add no behavior; one-instance interfaces; single-call factories; config objects wrapping a single primitive; plugin/registry patterns with only one registered implementation and no second on the horizon
3. Under-abstraction — Identical or near-identical logic copy-pasted across files with slight variation; repeated stream-parsing patterns across providers; duplicated kernel dispatch logic
4. Dead or disabled systems — Feature-flagged code that has no activation path; exported functions/types with zero callers in the monorepo; disabled strategies or routing paths; commented-out orchestration logic
5. Scope creep — Utility files that have grown into mini-frameworks; config objects that have accumulated behavior beyond pure data; packages whose index.ts exports exceed their documented purpose
6. Coupling hotspots — Files imported by 8+ other files; changes here force cascading updates; circular or near-circular dependencies

For each finding: file path, one-sentence problem description, fix direction, and risk (Low / Medium / High).

Agent 2: Documentation Accuracy

Compare every documented claim against what actually exists:

1. AGENTS.md architecture claims — Read each section describing package responsibilities, the ExecutionEngine loop phases, and build order. Does the code match? Flag drifted descriptions, removed phases, renamed exports
2. MEMORY.md accuracy — Check "Architecture (Post Refactor)" and "What Shipped" sections against current file contents. Flag entries describing files, patterns, or module layouts that no longer exist
3. Kernel extension docs — Do AGENTS.md instructions for "adding a new phase" / "adding a guard" / "adding a meta-tool" still work step-for-step against the current code?
4. Inline file headers — Are module-level JSDoc or block comments describing the right responsibility? Are @param / @returns still accurate?
5. apps/docs/src/content/docs/ — Do user-facing docs describe current public API behavior? Are there documented APIs (ReactiveAgentBuilder methods, strategy names, config fields) that no longer exist or have changed signature?
6. Known debt list — For each item in the Phase 1 debt table: has it been resolved without docs being updated? Update the debt table if so

For each finding: document file, specific claim, what's wrong in one sentence, correct description.

Agent 3: Simplification Opportunities

Find where complexity exceeds what the problem requires:

1. Configuration explosion — Config schemas with accumulated fields beyond any real use case; options that have never been set to non-default by any test or caller; YAGNI violations
2. Phase or strategy duplication — Multiple phases doing overlapping work (e.g., two context-assembly paths); strategies that share >70% of their logic and could be one strategy with a parameter
3. Indirection chains — A → B → C → D where B and C route without adding behavior; passes-through that exist only for historical reasons
4. Parallel systems — Two systems doing the same job (e.g., two tool-call parsers, two context builders, two error formatters); identify which is canonical and which is dead weight
5. Type gymnastics — Complex conditional types, deep infer chains, or mapped types where a simpler domain model would be clearer and equally safe
6. Premature extensibility — Registries, plugin hooks, or lifecycle systems built for N implementations but only 1 exists and no second is planned in the next milestone

For each finding: file path, description of unnecessary complexity, effort estimate (Low / Medium / High), impact estimate (Low / Medium / High).

---

Phase 3: Triage and Fix

Wait for all three agents. Categorize every finding:

Fix Immediately

Safe to fix now without architectural planning:

• Stale documentation claims in AGENTS.md, MEMORY.md, inline comments
• Dead exports (zero callers, confirmed unused)
• Incorrect or outdated kernel extension instructions
• Comment cleanup (narrating removed behavior, referencing deleted files)

Rule: fix is < 25 lines and touches ≤ 2 files. Anything larger → Flag for Planning.

Flag for Planning

Do NOT implement without a written plan:

• Dead code removal touching > 2 files (e.g., context-engine.ts cleanup)
• Collapsing parallel systems
• Restructuring config schemas
• Untangling layer violations

Add each to the Architecture Debt Register (see Output below).

Escalate to User

Needs explicit decision before any action:

• Parallel systems where the canonical choice is ambiguous
• Simplifications that change the public API of @reactive-agents/*
• Findings that would span > 3 packages
• Anything where fixing reveals a deeper design question

---

Output

When complete, report:

1. Documentation fixes applied — count and list of files changed
2. Top 3 simplification opportunities — file, problem, effort/impact
3. Architecture debt register update — append findings to the debt table in AGENTS.md under an ## Architecture Debt section (create if absent), formatted as:

| Area      | File              | Problem                    | Effort | Impact | Status |
| --------- | ----------------- | -------------------------- | ------ | ------ | ------ |
| Dead code | context-engine.ts | ~690 LOC dead behind flags | High   | High   | Open   |

4. Escalations — anything requiring explicit user decision

Keep the summary under 12 lines. The debt register carries the full detail.

---

Architecture Quick Reference

Pattern	Location	What it IS
Phase[] pipeline	kernel-runner.ts	Sequential kernel phases; (state, ctx) => Effect<KernelState>
Guard[] chain	phases/guard.ts	Tool-call safety checks; any guard can block
MetaToolHandler registry	phases/act.ts	Inline meta-tool dispatch; new tools = one registry entry
makeKernel() factory	react-kernel.ts	Custom kernel configurations; override default phase set
Two independent records	kernel-state.ts	messages[] (LLM sees) vs steps[] (systems observe)
Provider adapter hooks	provider-adapters/	4-hook adapter system, all wired: continuationHint, errorRecovery, synthesisPrompt, qualityCheck (+ parseToolCalls)
ExecutionEngine loop	packages/runtime/	10 phases: BOOTSTRAP → GUARDRAIL → STRATEGY-SELECT → THINK → ACT → OBSERVE → MEMORY-FLUSH → VERIFY → AUDIT → COMPLETE

Layer Boundary Rules

core → llm-provider / memory / tools / identity / observability / interaction
     → reasoning / guardrails / verification / cost / eval / a2a / gateway / orchestration / prompts
     → runtime
     → reactive-agents (facade)

• No upward imports (reasoning must not import runtime)
• No skip-layer imports (reasoning must not import reactive-agents)
• packages/reasoning may import: core, llm-provider, memory, tools only

Anti-Patterns to Flag on Sight

Anti-pattern	Known location	Signal
Dead code behind feature flags	context-engine.ts	buildDynamicContext, buildStaticContext
Untyped meta bag with as any	kernel-state.ts	meta: Record<string, unknown>
Disabled strategy routing	strategy-registry.ts	Local model routing commented out
Two-path context building	context-engine.ts + phases/context-builder.ts	Which is canonical?