plan

/ac:plan

Interactive planner that runs entirely on the main thread (Opus 5). Spawns read-only subagents for parallel research, reads referenced code itself, walks the user through every load-bearing decision via AskUserQuestion, audits the plan for reuse and quality before writing, then writes a tier-assigned plan to .ac/plans/<slug>/plan.md.

Request: $ARGUMENTS

Standing rules

These hold for the whole run, including after a compaction. Everything below this block is procedure; these are the bounds. They sit here because a re-attached skill keeps only its first 5,000 tokens after compaction (https://code.claude.com/docs/en/skills.md) and this body is far larger, so a rule further down is gone from context on exactly the long runs that need it.

Turn termination. Your turn ends on exactly one of: an AskUserQuestion call, the Stage 6 plan summary (or, under AUTO_MODE = true, the chained ac:execute run reaching its own terminal state), or a named BLOCKER from <auto_mode>. Nothing else ends it. Never end a turn by describing what you would do next.

Context. Auto-compaction summarizes older turns and the run continues. A filling context window is not a stopping condition and not a reason to defer work to a new session. When the procedure you need has been truncated away, re-invoke the ac:plan skill to restore this body and read .ac/plans/<slug>/checkpoint.json for where the run was.

Loop bounds come from disk, never from working memory. A counter you hold in context drifts across a long run; a file does not, and neither does a shell command's answer. Stage 5.5 reads its iteration number, its previous issue count, and both its gate verdicts out of LOG_PATH via the review-counters call in 5.5c. Run them and read the result; do not carry the numbers forward in your head and do not evaluate the comparisons yourself.

Progress surface. Call TaskList before creating any task, so a resumed session extends its own list instead of duplicating it. One task per stage, never one per decision.