Plan Execution

Execute plans by detecting the right mode from context, dispatching agents, and verifying results.

Mode Detection

Mode	When to Use	Details
Subagent (default)	Plan directory with manifest.json, independent tasks, current session	Fresh agent per task, two-stage review. See references/execution-modes.md
Batch	User wants separate session, or plan has 10+ tasks	Checkpoint-based execution, 3 tasks per batch. See references/execution-modes.md
Parallel dispatch	2+ independent tasks, no formal plan needed (e.g., multiple unrelated test failures)	One agent per problem domain, no plan required. See references/execution-modes.md
Team	User says "team", tasks need ongoing collaboration	Named agents, wave-based coordination. See references/execution-modes.md

Tier	Model	Use For
Heavy	Opus	Implementation, complex debugging, architecture
Medium	Sonnet	Code review, analysis, planning
Light	Haiku	Formatting, simple lookups, validation checks

Max 4 concurrent agents to avoid resource contention
Explicit file ownership per agent -- no two agents modify the same file
Agents read briefing files from disk -- never paste full briefing content inline
Re-read manifest.json at wave boundaries to recover state after context compaction
Include return format constraint in every spawn prompt (token budget + structure)
Stop and ask when blocked -- don't guess. See references/guardrails.md

references/execution-modes.md -- detailed process for each mode
references/orchestration-patterns.md -- agent spawn rules, templates, common mistakes
references/guardrails.md -- stop conditions, cleanup, stale plan detection

Required skills: git-workflow (worktrees, stack), writing-plans (creates plans this skill executes)