Architecture Design Skill

Trigger

• Keywords: architecture, architecture design, architecture doc, component diagram, 3-architecture, system design, document architecture

When NOT to Use

• Tech spec writing (use /tech-spec)
• Code implementation (use /feature-dev)
• Architecture consulting only (use /codex-architect)
• Implementation roadmap (use /deep-analyze)

Usage

/architecture                          # Auto-detect feature, create/update
/architecture <feature-keyword>        # Specify feature
/architecture --skip-debate            # Skip Phase 3 adversarial debate

Workflow

sequenceDiagram
    participant U as User
    participant C as Claude
    participant E as Explore Agent
    participant X as Codex
    participant D as Architecture Designer
    participant B as /codex-brainstorm

    C->>C: Phase 0: Context Resolution
    par Phase 1: Research
        C->>E: Track A: Code pattern analysis (background)
        C->>C: Track B: Read tech-spec (inline)
    end
    E-->>C: Component list + dependencies
    C->>X: Track C: Architecture advice
    X-->>C: Independent recommendations
    C->>D: Phase 2: Architecture Design
    D-->>C: Component + flow + decisions
    C->>B: Phase 3: Verification (debate)
    B-->>C: Equilibrium conclusion
    C->>C: Phase 4: Write 3-architecture.md
    C->>U: Auto-trigger /codex-review-doc

Phase 0: Context Resolution

Detect the target feature using the 5-level cascade.

See @skills/create-request/references/feature-context-resolution.md for the full algorithm.

# The wrapper, not the CLI directly: it owns the failure payload, so the full shape with
# `scan_error: true` arrives however the CLI fails — nonzero exit, signal, partial write, or a
# payload that is not the agreed shape. (Not when `node` itself is unavailable: nothing running
# under node survives that.) `|| echo '{}'` would produce a payload the gate cannot see as failure.
node scripts/resolve-feature.js

scan_error gate. Gate on scan_error !== false, not on scan_error === true. When it is not exactly false the four source sets are unknown, not empty — the corpus could not be enumerated (unreadable directory, broken taxonomy, no repository), or the resolver never ran and a shell fallback supplied a payload with no such field at all. {} is the shape that made the stricter test useless: it has no scan_error, so === true is false and the gate passes a payload that contains nothing. Do not proceed as though the feature has no authority documents — report and take the ⚠️ Need Human exit. A key may still be present, so a non-null key is not evidence the sets are complete.

"A tech spec resolves" means design_records holds an entry of type: tech-spec — the same resolution Track B uses.

design_records is an array, and more than one entry can be a tech spec, so "the entry" needs a rule rather than an assumption. docs/features/auto-loop-evolution/ is the live case: a split spec contributes 2-tech-spec/2-tech-spec.md and its sub-document 2-tech-spec/1-phase-d-hook-hardening.md, both type: tech-spec design records. Select in this order, and stop at the first that answers:

Filter first, then choose — every later rule reads the filtered list, never the whole set. Candidates are the design_records entries whose type is tech-spec; a requirements or architecture record is not a candidate at any step, and a rule phrased over "entries" rather than over candidates will select one. docs/features/codex-review-spec/ and docs/features/harness-engineering-rebrand/ are the live proof: neither has a tech-spec design record, each has exactly one canonical requirements record, and a canonicality test applied to the unfiltered set picks it.

Rows 1 and 4 are different answers and must not be collapsed: "there is no spec" is a fact the skill acts on, "there are two and I cannot tell which" is an ambiguity it must not resolve by picking. The set decides, and the set also names the file. canonical_docs is role-blind: it selects the tech spec from doc_inventory whatever role that document resolves to, so a spec that has declared itself History record or Work record is still non-null there while being absent from design_records. Reading the alias as evidence of design authority is exactly the confusion the source sets replace — and it is no better as a path selector: it is chosen across the whole inventory by type and canonicality, so a historical canonical 2-tech-spec.md beside a design-record variant 2-tech-spec-v2.md makes the set and the alias name different files. Take both the decision and the path from the design_records entry's own file; do not rejoin through the alias for either. Testing for the literal filename 2-tech-spec.md would read a split spec (2-tech-spec/2-tech-spec.md) or a variant (2-tech-spec-v2.md) as "no tech spec" and silently drop to code-only mode.

Scope Gate

For small features (tech-spec WBS has only 1 task, or no tech-spec and < 3 related files), suggest keeping architecture in tech-spec Section 3 instead of creating a separate document. Use AskUserQuestion to confirm.

Phase 1: Architecture Research (parallel)

Launch research tracks. Tracks A and B run in parallel; Track C runs after both complete.

Track A: Code Pattern Analysis (background)

Agent({
  description: "Analyze architecture patterns for <feature>",
  subagent_type: "Explore",
  run_in_background: true,
  prompt: "Analyze the codebase architecture for feature <key>:
    1. Trace execution paths of related modules
    2. Map component dependencies (imports, calls)
    3. Identify integration points with other features
    4. Read docs/architecture.md for global context
    Output: component list + dependency graph + integration points"
})

Fallback: subagent_type: "general-purpose" if Explore unavailable.

Track B: Tech-spec Extraction (inline)

The tech spec is a design record (design_records in the resolver output), which is exactly what this track wants: the intent, not the current behaviour. Track A supplies what the code actually does, and where the two disagree the code wins and the disagreement is worth stating in the architecture doc.

Resolve the file from design_records rather than assuming the name — a split spec lives at 2-tech-spec/2-tech-spec.md and a variant may be 2-tech-spec-v2.md, both of which the resolver classifies and a hard-coded filename misses. Use that entry's own file — do not rejoin through canonical_docs. The alias is selected independently from the whole inventory by type and canonicality (scripts/lib/doc-classifier.js § pickCanonicalDocs), so with a historical canonical 2-tech-spec.md beside a design-record variant 2-tech-spec-v2.md the set returns the variant and the alias returns the historical file. Filtering by role and then resolving the path through the alias silently swaps one for the other.

If a tech spec is present:

• Read Section 3 (Technical Solution) — architecture diagram, data model, API design
• Read Section 4 (Risks) — constraints, dependencies
• Read Section 7 (Open Questions) — unresolved design decisions

If no tech-spec: skip (code-only mode).

Track C: Codex Architecture Advice (after A+B)

Dispatch references/codex-prompt.md per @skills/codex-code-review/references/codex-transport.md § Start. The transport pins the sandbox and approval policy, so nothing is chosen here.

Provide feature context metadata only — never feed Claude's conclusions (per @rules/codex-invocation.md).

Save threadId for potential follow-up — a continuation goes through that reference's § Resume.

Graceful degradation is per outcome, subordinate to @skills/codex-code-review/references/codex-transport.md § Completion state machine — a single "Codex unavailable" rule flattened four different states into one action:

Phase 2: Architecture Design

Dispatch architecture-designer agent with merged research results:

Agent({
  description: "Design architecture for <feature>",
  subagent_type: "architecture-designer",
  prompt: `Design the architecture for <feature>.

  ## Input Context
  ${TECH_SPEC_SUMMARY}
  ${CODE_ANALYSIS}
  ${CODEX_ADVICE}

  ## Required Output
  Follow the output template at @skills/architecture/references/template.md:
  1. Component diagram (Mermaid flowchart)
  2. Component responsibility table
  3. Data flow (Mermaid sequence diagram)
  4. Integration points with existing systems
  5. Architecture decisions (AD-N: context → options → decision → rationale)
  6. Deployment considerations (if applicable)

  ## Constraints
  - Follow @rules/docs-writing.md conventions
  - Reference actual code (file:line, not invented)
  - Mark assumptions explicitly
  - Redact credentials/secrets per @rules/security.md`
})

Fallback: if architecture-designer agent unavailable, use solution-architect agent.

Phase 3: Verification (conditional)

Invoke /codex-brainstorm via Skill tool:

Skill("codex-brainstorm", `Evaluate the proposed architecture for <feature>.

Focus: scalability, maintainability, integration complexity, testability.

Constraints:
- Component diagram from Phase 2
- Tech-spec constraints from Phase 1
- Known risks`)

Must produce: threadId + equilibrium conclusion.

Skip Conditions

Graceful degradation: /codex-brainstorm timeout → record timeout in Verification section, still output document.

Phase 4: Output

Write docs/features/<key>/3-architecture.md using the output template.

See references/template.md for the full template.

Cross-References

Auto-insert links:

• > **Source**: [Tech Spec](./<design_records tech-spec entry .file>) — that entry's own file, never canonical_docs, and not the literal 2-tech-spec.md; a split spec lives one directory deeper and the hard-coded link is dead
• > **Request**: [Request](./requests/YYYY-MM-DD-*.md) (if active request found)

Auto-Trigger

After Write completes, auto-trigger /codex-review-doc per @rules/auto-loop.md.

Arguments

Verification

• Feature context resolved (create/update mode determined)
• Code research completed; the resolved tech spec read when one exists (§ Phase 0 says what "resolved" means, and a feature with no spec has none to read)
• Codex advice integrated — or its absence recorded as the exit 1 degradation of § Phase 1 Track C, which is this skill's only path to a design without it. There is no flag that skips Track C: --skip-debate skips Phase 3, and the no-tech-spec mode still dispatches Track C with (none — do not read a spec). Integrated or recorded, never neither — an unrecorded absence is indistinguishable from a dispatch nobody made
• Architecture design includes all required sections
• Mermaid diagrams are valid
• Architecture decisions use AD-N format with rationale
• Cross-references to tech-spec and request docs included
• /codex-review-doc passed (auto-triggered)
• No git add/commit/push executed

References

• references/template.md — Output template for 3-architecture.md
• references/codex-prompt.md — Codex independent architecture research prompt
• @skills/create-request/references/feature-context-resolution.md — 5-level feature detection

Examples

Input: /architecture
Action: Auto-detect feature → research (code + spec + Codex) → design → debate → write 3-architecture.md → /codex-review-doc

Input: /architecture statusline-config
Action: Resolve "statusline-config" → read tech-spec → research code → design → write → review

Input: /architecture --skip-debate
Action: Auto-detect → research → design → skip debate → write → review