Dynamic Workflow

Do not spawn agents for work a deterministic script/tool can do. Do not fan out small serial tasks.

For a large task:

1. classify each unit as deterministic, LLM judgment, or visual judgment;
2. build an explicit dependency DAG and declare read/write file scope;
3. use one shared model across fresh worker sessions unless the user explicitly requests otherwise;
4. serialize resource conflicts and isolate independent writers in worktrees;
5. bound/adapt concurrency instead of spawning as many agents as possible;
6. integrate one verified result at a time and immediately unlock newly ready dependencies;
7. keep large intermediate results on disk/evidence references instead of relaying them through parent chat context.

Read workflow.md for campaign and recovery rules.