gen-plan

Arguments: $ARGUMENTS

Load the run protocol first (Skill tool → baton:protocol; if that fails, Read ${CLAUDE_PLUGIN_ROOT}/skills/protocol/SKILL.md) and follow it. This is a doc-only flow: no run branch, no commits, and it may run on a dirty tree. This file only adds what is specific to gen-plan.

Idea

The side that writes remembers; the side that reads does not. The planner keeps its session so revisions build on each other. The analyst starts fresh every round so it judges the plan as a stranger would.

Flags

• INPUT: a path to an idea file (for example from /baton:gen-idea), or the idea as text.
• --out <path>: default docs/plan.md.
• --rounds N: maximum analyst rounds, default 3.
• --force: allow overwriting an existing output file.

Roles

Round

1. Planner. Round 1 (Agent call with name):

   BATON gen-plan · turn 1/<rounds> · role planner · run <run-id>
   
   INPUT:
   <the idea file's path and "read it first", or the idea text verbatim>
   
   Write a plan for this repository to <out>. Write no code and change no other file. Use these sections:
   - Goal
   - Context: the facts about this repository the plan relies on, with file references
   - Scope and non-goals
   - Steps: numbered; each with an acceptance check that a reviewer can run or read
   - Risks
   - Test plan
   - Decisions: each line "RESOLVED: <decision>" or "PENDING: <question> (options: ...)". Use PENDING only for what truly needs a human.
   You keep this session. A fresh analyst will review the plan and its objections will come back to you.

   Round n > 1 (SendMessage to <run-id>-planner): the header line, then "Revise <out> to answer these objections. Where you disagree, say why in the plan itself:" followed by the analyst's findings and next_prompt, verbatim.
2. Analyst (fresh):

   BATON gen-plan · turn <n>/<rounds> · role analyst · run <run-id>
   
   INPUT: <same as above>
   PLAN: <out>
   
   Review the plan against the INPUT and this repository. Object to anything wrong, missing, vague, untestable or risky; check that the Context matches the code. Return done: true only if you agree with the plan as written.
   Add the line "about_repo: yes|no": no if the plan isn't about this repository.

Stop

• about_repo: no: stop and say the plan doesn't fit this repository.
• done: true and no PENDING line in the plan: finished. Tell the user to read <out>, then run /baton:rlcr <out>.
• done: true with PENDING lines: stop and show those questions to the user. They decide, edit the plan, then run /baton:rlcr.
• --rounds used up without agreement: stop and show the last objections.
• A failed turn (protocol).