Architecture

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).