dev%3aplan（DEV 3aplan 搜索）

Compatibility: If AskUserQuestion is unavailable, present options as a numbered list and wait for the user's reply. If Task is unavailable, run parallel steps sequentially. The context: fork and agent: frontmatter fields are Claude Code-specific — on OpenCode and VS Code Copilot they are ignored and the skill runs inline using the current model.

Plan software architecture for a project or feature. Writes drafts to .codevoyant/plans/.

Step 0: Parse Args

Extract:

• FEATURE_NAME = first non-flag argument (slugify for filename)
• PLAN_MODE = value after --mode (arch or feat); if absent, leave unset — Step 1 will ask
• BG_MODE, SILENT as standard

If no FEATURE_NAME: ask "What are we designing architecture for?" (free text).

Derive:

• FEATURE_SLUG = lowercase, hyphens (e.g. "auth refresh flow" → "auth-refresh-flow")
• PLAN_DIR =.codevoyant/plans/{FEATURE_SLUG}

Create output directory:

mkdir -p "$PLAN_DIR/research"

Step 0.5: System Audit

git log --oneline -10
ls .codevoyant/plans/ 2>/dev/null || echo "(no existing plans)"
ls docs/architecture/ 2>/dev/null || echo "(no existing arch docs)"

If docs/architecture/README.md exists, read it — use as context so this plan doesn't duplicate existing decisions.

Step 1: Gather Design Context

If PLAN_MODE is unset, ask:

AskUserQuestion: question: "What kind of plan is this?" header: "Plan mode" options: - label: "Architecture (--mode arch)" description: "System or component design — produces task breakdown with LOE and blocking relationships" - label: "Feature implementation (--mode feat)" description: "A specific feature within a known architecture — produces a design doc only"

Set PLAN_MODE from the answer before continuing.

Ask:

1. "What scope are we designing?" — options: New feature | Refactor existing system | Cross-cutting concern (auth, logging, etc.) | Greenfield project
2. "What do you know about the design already?" — free text; may be empty ("I need you to research and propose")
3. "Confidence level?" — Already decided (document it) | Exploring options | Spike needed (too many unknowns)

Step 2: Parallel Research

Launch two background agents (model: claude-haiku-4-5-20251001, run_in_background: true):

Agent R1 — Codebase Scan: Glob/Grep for files, patterns, and systems relevant to this feature. Note existing architecture decisions, naming conventions, and test coverage. Return structured findings.

Agent R2 — Existing Architecture Docs: Read all files in docs/architecture/. Identify: what the current system looks like, what decisions are already recorded, what this feature touches. Flag gaps or contradictions with proposed design.

Wait for both agents. Synthesize: highlight what's new vs. what integrates with existing decisions.

Step 3: Architecture Design

Based on context and research, produce the architecture document content:

Sections (required):

Context — what system/feature this is in, why now

Design Decision — the architectural choice made (or top 2-3 options if exploring). For each option: trade-offs, complexity, reversibility.

Data Model — entities, relationships, storage. ASCII diagram if schema is non-trivial.

System Boundaries — what this feature owns vs. delegates. Use a Mermaid flowchart:

flowchart LR
    A[This feature] --> B[Dependency A]
    A --> C[Storage / Queue]

API Surface — new or modified interfaces (method, path/name, request/response shape). Mark N/A if internal only.

Key Decisions — table of one-way vs two-way doors:

Failure Modes — top 3 ways this can fail, with mitigation:

Open Questions — unknowns that need resolution before implementation starts

Out of Scope — explicitly deferred design concerns

Design principles:

• Boring by Default: name any existing library or pattern that could be reused instead of building
• If a section is unknown: write [spike needed], not omit it
• Decisions table: every one-way door must have a rationale

Step 4: Confirmation

Show a one-paragraph summary of the design. AskUserQuestion:

question: "Does this architecture look right?"
header: "Design Review"
options:
  - label: "Looks good — write the docs"
    description: "Write feature doc and update overview"
  - label: "Revise the design"
    description: "I'll describe what to change"
  - label: "Mark as exploratory"
    description: "Write as a proposal, not a decision"

Loop on revisions until "Looks good" or "Mark as exploratory".

Step 5: Write Plan Files

Write {PLAN_DIR}/plan.md with all architecture doc sections:

• Context
• Design Decision
• Data Model
• System Boundaries (Mermaid diagram preferred over ASCII)
• API Surface
• Key Decisions table
• Failure Modes table
• Open Questions
• Out of Scope

If "Mark as exploratory": prepend > **Status: Proposal** — not yet decided.

Task Breakdown (if PLAN_MODE is arch)

If PLAN_MODE is arch (or scope chosen was "Refactor existing system" or "Greenfield project"), append a ## Task Breakdown section to plan.md:

For each implementation task identified during design, write a self-contained entry rich enough for an autonomous agent to run /spec:new and /spec:bg without further human input:

{task name}

• LOE: {N} hours (rough estimate)
• Blocks: {list of task names this task must complete before, or "none"}
• Blocked by: {list of task names that must complete first, or "none"}
• Architecture reference: {PLAN_DIR}/plan.md → ## {Section name} (e.g. ## Design Decision, ## API Surface)
• Scope: {one paragraph — what specifically must be built or changed to implement this task}
• Key constraints: {relevant ONE-WAY door decisions from the Key Decisions table that apply to this task}
• Acceptance criteria:

- {specific, verifiable condition checkable in under 5 minutes} - {another AC — e.g. unit test passes, endpoint returns expected shape, migration is idempotent}

Include a Mermaid dependency graph if there are blocking relationships:

graph TD
  A[Task A] --> B[Task B]
  A --> C[Task C]
  C --> D[Task D]

Step 7: Report + Notify

Register the plan:

npx @codevoyant/agent-kit plans register \
  --name "{FEATURE_SLUG}" \
  --plugin dev \
  --description "{first line of Context section}" \
  --total "{task count from Task Breakdown, or 0}"

Report:

Plan written to .codevoyant/plans/{FEATURE_SLUG}/plan.md
Run /dev:approve to promote to docs/architecture/.

If BG_MODE=true and SILENT=false:

npx @codevoyant/agent-kit notify --title "dev:plan complete" --message "Architecture plan saved: .codevoyant/plans/{FEATURE_SLUG}/"