subagent-development

Subagent Development

$ARGUMENTS

Execute implementation plans by dispatching fresh subagents per task, then running a two-stage review gate: spec compliance first, code quality second. Fresh context per subagent prevents accumulated confusion. Two-stage review catches different failure modes: spec review catches wrong behavior, quality review catches bad structure.

Usage

/subagent-development [plan file or task description]

Why This Works

Property	Benefit
Fresh subagent per task	No accumulated context drift or confusion
Spec review first	Catches wrong behavior before quality review wastes time on wrong code
Quality review second	Catches structural issues after behavior is confirmed correct
Sequential tasks	No merge conflicts, each task builds on verified previous work

Process Flow

Read plan
    |
    v
Extract ordered task list
    |
    v
For each task:
    |
    +---> [1] Dispatch IMPLEMENTER subagent
    |         |
    |         v
    |     Handle status (see Status Protocol)
    |         |
    |         v
    +---> [2] Dispatch SPEC REVIEWER subagent
    |         |
    |         v
    |     APPROVED? --no--> fix issues, re-review
    |         |
    |        yes
    |         |
    |         v
    +---> [3] Dispatch QUALITY REVIEWER subagent
    |         |
    |         v
    |     Critical issues? --yes--> fix, re-review
    |         |
    |         no
    |         v
    |     Mark task complete
    |
    v
Next task (or done)

Step 1 -- Read and Parse Plan

Read the plan file. Extract:

1. Ordered task list -- each task with description, acceptance criteria, file scope
2. Global constraints -- what must NOT change, architecture rules, shared conventions
3. Dependencies -- which tasks depend on which (execute in dependency order)

Present the task list to the user. Wait for approval before proceeding.

Step 2 -- Execute Tasks Sequentially

Only in Claude Code, apply the model-routing-patterns skill when choosing executors or creating agent definitions. Delegate to codex:codex-rescue only when its plugin is installed, enabled and callable in this session. Otherwise use the installed native agents and their configured models. A context without the Agent tool returns the dispatch decision to its supervisor; it does not invent a tool or bypass the client. Preserve explicit user choices and verify actual completion before accepting a delegated result.