plan

Plan

You are pair-planning with the user. Explore the code to build context, ask the user decisions only they can make, and write findings into the plan file as you go. The plan file is the ONLY file you may edit during planning: it starts as a rough skeleton and gradually becomes the final plan.

Scope: main agent handles the interview and the writes. The built-in Explore subagent is allowed for read-only research when a single directed search is clearly insufficient (broad codebase survey, architectural mapping, or tasks you can already predict will need more than 3 queries). Prefer inline Read / Glob / Grep for narrow, directed scans. Never spawn other custom subagents, never use Agent(Plan). Trivial tasks bypass the plan file entirely and execute inline.

Phase 1: Initialize

1. Parse $ARGUMENTS. If empty, ask the intent first:

   {
     "questions": [{
       "question": "What should I plan?",
       "header": "Request",
       "multiSelect": false,
       "options": [
         {"label": "Build a new feature", "description": "Create something from scratch."},
         {"label": "Refactor existing code", "description": "Restructure an area without changing behavior."},
         {"label": "Fix a bug", "description": "Reproduce, diagnose, fix with a regression test."},
         {"label": "Architecture or migration", "description": "Cross-cutting change, likely multi-phase."}
       ]
     }]
   }
   

   Follow up as free text for the specific description.

2. Continuation check: if $ARGUMENTS resolves to an existing .ac/plans/<slug>/ROADMAP.md, this is a Mode B continuation. Read the ROADMAP, pick the next phase with status: draft, jump to Phase 3 scoped to that phase only. Announce: Mode B continuation, planning phase-0N-<name>.

3. Ensure .ac/plans/ exists in cwd; mkdir -p .ac/plans if not.

4. Pre-invoke twin skills (silent no-op if missing):

   • Skill("my-coding") for Task code snippets.
   • Skill("my-language") for plan prose.