01ONE OR MANY

One flag picks one worker or many:

# linear pipeline (default):
kineti run --goal "Refactor database queries" --mode single

# parallel feature workers after ONE spec approval:
kineti run --goal "Build auth, billing, and an admin dashboard" \
    --mode swarm --cap 100

02SPLIT THE WORK

In swarm mode, steps 1–6 run once. The architecture file must list a work split — one line per worker with folder pattern and what it needs:

## Task Partition
- T1: database layer | scope: src/db/** ; migrations/** | deps: -
- T2: api routes     | scope: src/api/**             | deps: T1

If two workers share files or make a loop, it goes back to step 4 before any code is written. One y at the spec stop approves the whole split — one choice for all workers.

03KEEP WORK SEPARATE

Each worker does build → review → qa in its own folder. Workers run in groups, up to [execution].max_parallel_workers at once:

Each worker keeps its own log (journal.w-<task>.jsonl) with hashes. If one worker crashes, the others keep going.

04JOIN THE WORK

Joining follows fixed steps:

1. Merge workers one by one into the main folder.
2. If files clash: try one auto-fix.
3. If still clashing: stop and show the diff to you.

After joining, the combined code needs fresh tests before steps 10–13. Worker logs are joined to the main log with a two-parent record; kineti merge --branch <id> joins one branch. Orphan or late-extended branches block ship.