Update Architecture

Architecture Documentation Update Skill

When to Use

Use this skill when:

• User says "update architecture docs", "sync architecture docs", "update arch docs", "refresh architecture"
• When architecture documentation is out of sync with implementation

Target Selection (Required)

Update ONE location at a time - updating all locations simultaneously creates too much context. Identify which location to update based on user request.

If user doesn't specify, ask which component or area of the codebase to update using AskUserQuestion.

Critical Constraints

NEVER

• Read existing architecture docs BEFORE understanding the codebase (prevents context pollution)
• Catalog components without understanding their PURPOSE
• Document data flows without understanding WHY they exist that way
• List patterns without understanding what PROBLEM they solve
• Patch sections - rebuild from understanding instead
• Trust old documentation over understood principles
• Create diagrams that don't reflect understood architecture
• Document WHAT without explaining WHY
• Assume a WRITE operation implies a corresponding READ (verify both independently)
• Document data flows without tracing actual RETRIEVAL code

ALWAYS

• Focus on ONE target location at a time - ask user if not specified
• Understand before documenting - answer the 6 key questions (Purpose, Constraints, Design Rationale, Collaboration, Invariants, Data Lifecycle)
• Build understanding from code FIRST, then compare with existing docs
• Write documentation that explains WHY, not just WHAT
• Trace READS and WRITES separately - verify each data flow has both ends in actual code
• Establish authority by retrieval patterns - what code QUERIES determines source of truth
• Validate through understanding - errors violate understood principles
• Use /mermaid skill for diagrams that reflect understood architecture
• Maintain the standard 8-file structure per location
• Preserve valid structural elements (navigation, formatting) from existing docs

Standard 8-File Structure

Every architecture location contains these files:

Workflow

Core Principle: Understand, Don't Catalog

CRITICAL: The goal is to deeply UNDERSTAND the architecture - the WHY behind every design decision, not just WHAT exists. An agent that truly understands the architecture will never document incorrect patterns because they would violate the understood principles.

Do NOT:

• Read existing architecture docs before exploring code (prevents context pollution)
• Create checklists of things to look for (leads to shallow pattern matching)
• Document components without understanding their purpose
• Describe data flows without understanding why they exist that way
• Assume a WRITE implies a READ (verify each independently)
• Assume data in a location is used just because it exists there

Instead:

• Explore code to understand the PROBLEM being solved
• Understand the CONSTRAINTS and challenges the system faces
• Understand WHY each design decision was made
• Understand HOW components collaborate to achieve the goal
• Understand the INVARIANTS that must always hold true
• Trace READS and WRITES separately for every data flow
• Verify authority by finding RETRIEVAL code, not just storage code

Step 1: Git Diff Analysis (Single Location)

Launch 1 subagent to analyze code changes for the selected target location since the last architecture update:

Subagent 1 (Git Diff):
  - Find last commit: git log -1 --format="%H" -- {architecture_docs_path}/
  - Diff since then: git diff <commit>..HEAD -- {target_code_path}/
  - Report: new files, modified files, deleted files, summary of changes

Step 2: Deep Architecture Understanding (Single Location)

Launch parallel Explore subagents to build TRUE UNDERSTANDING of the target module. These subagents must NOT read any architecture/*.md files - only source code.

Each exploration subagent must answer these questions:

1. PURPOSE: What problem does this module/component solve?
   - What would break if this didn't exist?
   - What user need or system requirement does it address?

2. CONSTRAINTS: What challenges does this module face?
   - What are the reliability requirements?
   - What are the consistency requirements?
   - What failure modes must be handled?

3. DESIGN RATIONALE: Why was it built THIS way?
   - What alternatives were rejected and why?
   - What trade-offs were made?
   - What principles guide the design?

4. COLLABORATION: How do components work together?
   - What is the flow of data/control?
   - What are the boundaries between components?
   - What contracts exist between components?

5. INVARIANTS: What must ALWAYS be true?
   - What assumptions can never be violated?
   - What guarantees does this module provide?
   - What guarantees does it depend on from others?

6. DATA LIFECYCLE: What is the complete lifecycle of each data entity?
   - Where is this data CREATED?
   - Where is this data WRITTEN/PERSISTED?
   - Where is this data READ/RETRIEVED? (trace actual retrieval code)
   - Is this data EVER read back, or only written?
   - If multiple storage locations exist, which does the code QUERY?

Subagent exploration structure:

Subagent 2: Understand the module's CORE PURPOSE
  - Read entry points, main classes, public interfaces
  - Answer: What is this module FOR? What problem does it solve?
  - Answer: What would happen if this module didn't exist?

Subagent 3: Understand the module's DATA ARCHITECTURE
  - Trace how data enters, transforms, persists, and exits
  - Answer: Where is the source of truth? Why?
  - Answer: What consistency guarantees are provided? How?
  - Answer: What would break if data flowed differently?

  CRITICAL - Trace READS and WRITES separately:
  - For each storage location (database, files, memory), trace:
    * What code WRITES to this location?
    * What code READS from this location?
  - Don't assume writes imply reads - verify each independently
  - Authority is determined by RETRIEVAL patterns, not storage patterns
  - Ask: "When component X needs this data, what does it query?"

Subagent 4: Understand the module's CONTROL FLOW
  - Trace execution paths, state transitions, decision points
  - Answer: What orchestrates the workflow? Why that approach?
  - Answer: How are failures handled? Why that strategy?
  - Answer: What are the critical paths?

Subagent 5: Understand the module's BOUNDARIES
  - Identify interfaces with other modules
  - Answer: What does this module depend on? What depends on it?
  - Answer: What contracts exist at boundaries?
  - Answer: How is coupling minimized?

Subagent 6: Understand the module's EXTENSION POINTS
  - Identify patterns for customization, configuration
  - Answer: How can behavior be modified without code changes?
  - Answer: What is intentionally flexible vs fixed?

Subagent 7: Trace DATA LIFECYCLE (Critical for correct data flow documentation)
  - For each storage location found (database tables, files, memory structures):
    * Search for all WRITE operations to this location
    * Search for all READ operations from this location
    * If writes exist but NO reads found -> this is write-only (not part of data flow)
    * If reads exist, trace: what component reads? when? why?
  - Establish AUTHORITY:
    * When a component needs data, what does it query FIRST?
    * If data exists in multiple places, which location is consulted?
    * The code that RETRIEVES data reveals the true source of truth
  - Document the complete lifecycle:
    * Creation -> Transformation -> Persistence -> Retrieval -> Usage
    * Note any gaps (e.g., "persisted but never retrieved")

Each subagent produces an understanding document that explains the WHY, not just the WHAT.

CRITICAL: Data Flow Verification

Before documenting any data flow (A -> B), the explorer MUST verify:

1. There is code that WRITES from A
2. There is code that READS into B
3. These operations are connected (B actually uses what A writes)

A write operation observed does NOT prove a read exists. Trace both independently.

Step 3: Synthesize Understanding Into Documentation

Based on the deep understanding from Step 2, rebuild documentation that reflects TRUE understanding. Launch parallel subagents to write documentation from the understanding gained:

Each documentation rebuild subagent receives the UNDERSTANDING from Step 2
and writes documentation that explains:

Subagent N: Rebuild overview.md
  - Write from understanding of PURPOSE and DESIGN RATIONALE
  - Explain what the module does and WHY it exists
  - Create diagrams that show HOW components achieve the goal

Subagent N+1: Rebuild components.md
  - Write from understanding of COLLABORATION and BOUNDARIES
  - Explain each component's ROLE in achieving the purpose
  - Describe WHY each component exists, not just what it does

Subagent N+2: Rebuild data-models.md
  - Write from understanding of DATA ARCHITECTURE and INVARIANTS
  - Explain WHY data flows the way it does
  - Document the source of truth and WHY it's the source of truth
  - Explain what guarantees are provided and HOW

Subagent N+3: Rebuild workflows.md
  - Write from understanding of CONTROL FLOW and CONSTRAINTS
  - Explain WHY the workflow is structured this way
  - Document failure handling and WHY that strategy was chosen

Subagent N+4: Rebuild patterns.md
  - Write from understanding of DESIGN RATIONALE
  - Explain each pattern's PURPOSE - what problem it solves
  - Document WHY this pattern was chosen over alternatives

Subagent N+5: Rebuild configuration.md
  - Write from understanding of EXTENSION POINTS
  - Explain what each config controls and WHY it's configurable

Subagent N+6: Rebuild directory-structure.md
  - Write from understanding of BOUNDARIES and COLLABORATION
  - Explain WHY the code is organized this way

Key principle: Every statement in the rebuilt docs should be traceable to understood purpose or constraint.

Step 4: Dry Walkthrough Validation

Launch validation subagents to perform a dry walkthrough of the rebuilt documentation - tracing every documented relationship back to actual code:

4a. Diagram Walkthrough

For each diagram in the rebuilt documentation:

Subagent V1: Walkthrough each connection in diagrams

  For each arrow/connection (A -> B) in a diagram:
  1. LOCATE: Find the actual code where A connects to B
  2. CHARACTERIZE: What does this connection actually do?
  3. VALIDATE: Does the diagram accurately represent reality?
  4. REPORT: Document findings

4b. Data Flow Walkthrough (CRITICAL)

For each data flow described in documentation:

Subagent V2: Verify data flows by tracing READS and WRITES separately

  For each documented flow (Source -> Destination):

  1. VERIFY THE WRITE:
     - Find code that writes from Source
     - What function/method performs the write?

  2. VERIFY THE READ:
     - Find code that reads into Destination
     - Does this read ACTUALLY use what was written in step 1?

  3. CHECK FOR PHANTOM FLOWS:
     - If you find a WRITE but NO corresponding READ -> flow doesn't exist
     - A write operation alone does NOT prove data flows to another component

  4. ESTABLISH AUTHORITY:
     - If data exists in multiple locations, which does the code QUERY?
     - The retrieval path reveals the true source of truth

4c. Compare With Existing Docs

Subagent V3: Compare validated rebuild against existing docs

  After walkthrough validation, compare:
  - Connections in old doc that don't exist in code -> STALE
  - Connections in old doc with wrong direction/label -> INCORRECT
  - Connections missing from old doc but found in walkthrough -> MISSING
  - Produce report explaining each discrepancy with code evidence

Step 5: Apply Rebuilt Content

For each affected file in the target location:

1. Use rebuilt content as the source of truth - not the old doc
2. Preserve valid structural elements - navigation links, formatting conventions
3. Replace stale sections entirely - don't try to patch incorrect content
4. Use /mermaid skill for any diagram creation or updates
5. Update cross-references - ensure links to other docs remain valid
6. Report discrepancies found - log what was incorrect in old docs

Step 6: Final Validation

Launch parallel final validation subagents to ensure consistency within target location:

• Validate cross-references between the 8 files
• Validate diagrams match rebuilt content
• Validate content consistency across files
• Spot-check rebuilt content against actual code one final time

Related Skills

Mermaid Skill (/mermaid)

When architecture documentation requires new or updated Mermaid diagrams, invoke the /mermaid skill. This ensures diagrams are properly formatted and validated.

Output

This skill modifies architecture files for the selected target location. Changes and discrepancies are reported to terminal:

=== Architecture Documentation Update Report ===

Target: {target_location}

Understanding Phase:
- PURPOSE: {what problem the module solves}
- INVARIANTS: {what must always be true}
- KEY INSIGHTS: {important design rationale discovered}

Walkthrough Validation:
- {N} diagram connections traced
- {X} valid, {Y} issues found
- Each issue includes: code evidence, why it's wrong, action taken

Discrepancies Found:
- {file}: {what was wrong} -> {what code actually shows}

Updated Files:
- {list of files modified with action taken}

Final Verification: {pass/fail with details}

Why Understanding Over Cataloging?

Cataloging (what we avoid):

• Lists components without explaining why they exist
• Looks for specific keywords without understanding why they matter
• Compares diagram shapes without understanding what they represent

Understanding (what we do):

• "Why does this module exist? What problem does it solve?"
• "What would break if this component didn't exist?"
• "What constraints make this design necessary?"
• "What invariants must always hold true?"

An agent that truly understands the architecture will naturally produce correct documentation because incorrect statements would violate understood principles. The understanding itself prevents errors.

Why Rebuild Instead of Patch?

Patching existing documentation risks:

• Context pollution: Reading stale docs first biases toward preserving incorrect info
• Shallow updates: Patches fix surface issues while deeper misunderstandings persist
• Missing the why: Patches change WHAT without updating WHY

Rebuilding from understanding ensures:

• Principled documentation: Every statement follows from understood purpose/constraints
• Self-correcting: Errors are caught because they violate understood invariants
• Future-proof: Documentation reflects WHY, so it guides future changes correctly